From task-specific AI to multimodal foundation models in PET/CT: a readiness framework for clinical translation

1 Introduction

Artificial intelligence (AI) has entered nearly every stage of the PET imaging chain, including detector and acquisition optimization, image reconstruction, denoising, attenuation correction, lesion detection, segmentation, quantitative biomarker extraction, response assessment and outcome prediction. During the first wave of adoption, each model was generally optimized for one task and one narrowly defined data distribution. These systems can be highly effective, but their value often diminishes when tracer biology, scanner hardware, reconstruction settings, patient population or clinical question changes. Recent reviews of PET AI consistently identify external validation, standardization, and workflow integration as persistent bottlenecks (1–4).

Foundation models propose a different development strategy. Instead of learning each endpoint from scratch, a large model is pretrained on broad data using self-supervised, contrastive or generative objectives and subsequently adapted through prompting, lightweight fine-tuning, retrieval or task-specific heads (5). Medical vision-language models extend this strategy by aligning images with reports, electronic health records or other language supervision, thereby enabling zero-shot classification, cross-modal retrieval, visual question answering, and report generation. Large generalist biomedical models and three-dimensional CT vision-language models have demonstrated that diverse supervision can support multiple downstream tasks from a shared representation (6–9).

Nuclear medicine should not treat this transition as a simple transfer of methods from radiography or CT. PET is a quantitative, tracer-dependent measurement of biology. Uptake depends on injected activity, uptake time, reconstruction, motion, attenuation and scatter correction, partial-volume effects and the physiology of the radiopharmaceutical (1–3, 10). A model can produce a visually plausible image or fluent report while corrupting standardized uptake values (SUVs), suppressing small lesions, confusing physiologic and pathologic uptake, or inventing comparative findings. The central challenge is therefore to build models that align molecular signal, anatomy and clinical semantics without erasing the physical meaning of PET (11–13).

This focused narrative review has three aims. First, it maps the evolution from task-specific PET/CT AI to reusable and multimodal foundation models. Second, it critically evaluates emerging PET-native promptable and vision-language systems against the requirements of clinical nuclear medicine. Third, it proposes MOSAIC-PET, an author-developed conceptual framework for PET/CT foundation-model development, evaluation, and lifecycle governance. Existing PET-focused AI reviews have primarily examined task-specific applications and recurring challenges such as standardization, external validation, and workflow integration (1–4), whereas recent reviews of foundation models and vision-language systems have largely considered these developments across medical imaging more broadly (14–17). The distinctive contribution of the present review is to connect these two areas by examining foundation-model development through the quantitative, tracer-dependent, multimodal, and longitudinal requirements of molecular imaging. In addition, we provide an operational taxonomy that distinguishes emerging systems according to demonstrated capabilities and link these distinctions to PET/CT-specific translational evidence requirements. Because PET-native foundation, vision-language, and agentic systems have appeared in rapid succession during 2025–2026, a synthesis now is timely and may help establish an evaluation discipline before premature or overstated clinical claims become entrenched. The framework is intended to make the review actionable for nuclear medicine physicians, imaging scientists, AI developers, clinical trialists, and regulators. As illustrated in Figure 1, PET/CT AI evolves from narrowly optimized models toward reusable representations and multimodal foundation models capable of supporting segmentation, retrieval, quantitative analysis, reporting, and longitudinal reasoning.

Accordingly, this review addresses three research questions: (1) How has PET/CT artificial intelligence evolved from task-specific models toward reusable pretrained representations, promptable systems, multimodal and vision-language models, and agentic workflows, and what capabilities have been demonstrated to date? (2) What PET/CT-specific technical, quantitative, biologic, and clinical requirements constrain the generalizability and translation of these emerging systems? (3) What evidence is needed to evaluate the readiness of PET/CT foundation-model systems for clinical translation, and how can these requirements be organized within a PET-specific evaluation framework?

1.1 Review scope, literature approach, and eligibility criteria

To address these research questions, we conducted a focused narrative search of PubMed, arXiv, and official conference proceedings repositories for English-language literature published from January 2021 through July 11, 2026. Search concepts combined nuclear medicine, positron emission tomography, PET/CT, total-body PET, artificial intelligence, deep learning, self-supervised learning, foundation model, multimodal model, vision-language model, segmentation, reconstruction, report generation, and clinical translation. The January 2021 boundary was chosen to focus the synthesis on the contemporary foundation-model era, while earlier standards and foundational studies were retained when necessary for technical, methodological, or historical context.

Eligible sources included peer-reviewed original studies, multicenter evaluations, conference proceedings, benchmark and challenge reports, public-dataset studies, relevant review or consensus articles, regulatory and reporting guidance, and recent preprints when directly relevant to emerging PET/CT foundation-model development. Studies were included when they addressed PET or PET/CT artificial intelligence, pretrained or reusable representations, foundation models, multimodal or vision-language learning, reconstruction or enhancement, segmentation, prediction, report generation, agentic workflows, or clinical translation. General medical-imaging foundation-model studies were included selectively when they provided directly relevant methodological or translational context for PET/CT. Sources were excluded if they were non-English, fell outside the January 2021–July 11, 2026 study period without a necessary foundational or standards-related role, lacked direct relevance to PET or PET/CT or directly applicable medical-imaging foundation-model methodology, or fell outside the AI and translational scope of this review. Because PET/CT foundation-model research is evolving rapidly, several of the most recent and directly relevant developments available at the time of the search were still available only as preprints. These reports were explicitly identified as non-peer-reviewed and interpreted as preliminary; peer-reviewed evidence and consensus or regulatory guidance were given greater weight in clinical and translational conclusions, and preprints were not used alone to support claims of clinical readiness.

2 Why PET/CT requires a domain-specific foundation-model strategy

2.1 PET is a molecular measurement, not only an image

PET intensities have quantitative and biologic interpretations that are more constrained than ordinary natural-image pixels. SUV-based response criteria, metabolic tumor volume, total lesion glycolysis and kinetic parameters can influence staging, treatment response assessment and prognosis. Accordingly, AI enhancement must be evaluated not only for image similarity but also for lesion detectability, uptake bias, repeatability and clinically meaningful thresholds. Multicenter low-count studies have shown that deep-learning enhancement can preserve diagnostic confidence and SUV agreement, while also underscoring the need for reader assessment and quantitative agreement appropriate to the intended use (10, 11, 18).

The same principle applies to generative and diffusion models. A denoiser may improve signal-to-noise ratio while changing lesion contrast, smoothing intratumoral heterogeneity or hallucinating uptake. Recent three-dimensional diffusion work has improved robustness across scanners, tracers and dose levels, yet clinical adoption still requires lesion-level error analysis and prospective comparison with the acquired reference data (19–21).

2.2 Hybrid imaging creates paired but asymmetric modalities

PET and CT are spatially linked but semantically asymmetric. CT provides high-resolution anatomy and attenuation information, whereas PET reveals tracer-specific physiology at lower spatial resolution. Naive channel concatenation, late fusion or independent encoders may underuse this complementarity or permit one modality to dominate. Cross-modal PET/CT models have improved prediction over single-modality approaches in specific clinical settings, including occult nodal metastasis in lung cancer, while self-supervised joint PET/CT denoising demonstrates the value of paired learning without fully aligned normal-dose targets (22, 23).

Hybrid acquisition also introduces failure modes that are not present in single-modality models: respiratory or patient motion can misregister uptake and anatomy; metallic implants and contrast can alter CT appearance; PET field of view and CT protocol may differ; and the CT component may be low dose or non-diagnostic (24, 25). PET/CT foundation models must therefore learn correspondence while retaining the ability to detect and communicate misalignment.

2.3 Tracer diversity is both an opportunity and a source of domain shift

FDG dominates existing AI datasets, but nuclear medicine spans PSMA-targeted agents, somatostatin-receptor tracers, amyloid and tau agents, amino-acid tracers, cardiac perfusion tracers and an expanding theranostic ecosystem. A representation that generalizes across tracers could enable label-efficient learning and rare-indication transfer. Conversely, anatomic uptake that is abnormal for one tracer can be expected for another. Reviews of PSMA PET/CT and total-body PET emphasize that tracer-specific biology and acquisition protocols cannot be treated as nuisance variation (26–28). Table 1 summarizes the principal differences between task-specific PET/CT AI and multimodal foundation models in terms of training objectives, supported modalities, adaptation strategies, outputs, and translational risks.

Table 1. Conceptual differences between task-specific AI and multimodal PET/CT foundation models.

3 From task-specific AI to foundation models in PET/CT

3.1 Reconstruction, denoising and attenuation correction

Reconstruction is one of the most mature PET AI applications. Deep networks can learn image-space denoising, direct or unrolled reconstruction, attenuation and scatter correction, and mappings between low-count and standard-count acquisitions. Recent works show a shift from convolutional post-processing toward physics-informed, self-supervised, diffusion and prompt-guided methods (2, 12, 29, 30). Foundation-model development is also extending to raw PET measurements. ALL-PET introduced a latent-diffusion model operating directly on projection-domain data (sinograms) and demonstrated reuse across low-dose reconstruction, attenuation correction, delayed-frame prediction, and tracer separation. Although the study illustrates reusable pretraining across PET image-formation tasks, it remains a preprint and requires independent multicenter clinical validation (31).

Potential clinical benefits include lower injected activity, shorter acquisition, improved throughput, and potentially more tolerable scanning. However, traditional perceptual metrics such as peak signal-to-noise ratio or structural similarity are inadequate if SUV bias, lesion conspicuity or response classification changes. Prospective, multivendor reader evidence remains much less common than retrospective technical validation (11, 18). Synthetic CT and CT-free attenuation correction add another layer of risk because errors can propagate into PET quantification. Recent studies of synthetic CT in PET and radiotherapy emphasize that evaluation must include attenuation maps, reconstructed PET bias and clinically relevant regional analyses, not only synthetic-image appearance (24, 25).

3.2 Detection, segmentation and whole-body tumor burden

Whole-body lesion segmentation is the bridge between task-specific models and foundation models. Public datasets and challenges have accelerated progress. The FDG-PET/CT-Lesions collection provided 1,014 studies with expert tumor annotations, while HECKTOR established reproducible PET/CT benchmarks for head-and-neck primary tumor and nodal segmentation and outcome prediction (32–34).

The autoPET challenge showed that accurate whole-body lesion segmentation is feasible and highlighted the importance of data quality and quantity alongside architectural design. Yet false-positive physiologic uptake, small lesions, diffuse marrow or splenic disease, post-treatment inflammation and uncommon cancers remain difficult. Challenge performance should therefore be interpreted as a benchmark under defined conditions, not evidence of universal clinical readiness (35, 36). Broader task-specific systems have begun to span diseases and tracers. A semi-supervised transfer-learning model evaluated 611 FDG PET/CT and 408 PSMA PET/CT scans, performed segmentation across several malignancies and derived prognostic measures for multiple downstream endpoints. This is an important step toward reusable representations, but the architecture still optimizes a defined family of tasks and depends on curated cohorts (37).

Federated learning offers one route to multi-institutional model development without centralizing images. In a six-center head-and-neck PET/CT study, federated transformer models approached centralized segmentation performance. Such approaches reduce data movement but do not automatically solve harmonization, annotation inconsistency, privacy leakage, governance or performance disparities across sites (38).

3.3 Diagnosis, response assessment and prognosis

Radiomics and deep learning have been applied to tumor characterization, nodal staging, therapy response and survival. More rigorous studies combine imaging with clinical covariates, use independent or prospective cohorts and evaluate added value beyond established comparators. For example, a PET/CT cross-modal signature for occult nodal metastasis in clinical N0 non-small cell lung cancer was validated in internal, external and prospective cohorts and outperformed single-modality models and physicians for the studied endpoint (22).

Nevertheless, prediction models remain vulnerable to selection bias, treatment-confounder effects, inconsistent endpoints and optimism from repeated feature selection. Radiomic features are sensitive to segmentation, reconstruction, voxel size and harmonization. Further work therefore calls for standardized acquisition, transparent preprocessing, calibration, external validation and clinically meaningful decision analysis (39, 40).

3.4 Reusable representations and foundation-model pretraining

The technical transition is enabled by three ideas. First, self-supervised learning extracts supervisory signals from the images themselves through masking, reconstruction, contrastive learning or temporal consistency (8, 14, 15). Second, language provides scalable semantic supervision because reports contain anatomy, findings, comparisons and diagnostic impressions (6, 7, 16, 17). Third, parameter-efficient adaptation allows a shared encoder to support new tasks with fewer labeled examples (5, 16, 17).

Generic computer-vision milestones shaped medical development. Contrastive language-image pretraining aligned visual and text embeddings at scale, while the Segment Anything Model introduced promptable segmentation. Medical adaptations show both the promise and limits of transfer. MedSAM was trained on more than 1.5 million image-mask pairs across 10 modalities and demonstrated broad segmentation capability; however, independent evaluations of generic SAM found reduced performance for low-contrast or poorly bounded medical targets (41–44).

Recent evidence also emphasizes the importance of modality-aligned pretraining. Dual-DINOv3 independently pretrained PET- and CT-specific self-supervised encoders and fused their representations through multiscale cross-attention for PET/CT tumor segmentation. Evaluation on the HECKTOR benchmark showed that natural-image initialization degraded performance relative to training from scratch, whereas domain-aligned CT pretraining improved segmentation and dual-stream PET/CT pretraining achieved the best overall performance. Although the authors describe the framework as a PET/CT foundation model, only one downstream task was evaluated; under the operational taxonomy proposed in Table 2, it is therefore classified conservatively as a task-specific segmentation model using reusable modality-aligned representations (45).

Table 2. Operational taxonomy of PET/CT foundation-model systems. Classes are assigned according to demonstrated capabilities rather than model size.

The classes are capability categories rather than ordinal quality levels. F2 and F3 are nonsequential, and the agentic category is orthogonal because an agentic system may orchestrate task-specific tools, foundation models, or both. Capability class should be reported separately from evidence maturity and clinical readiness.

Vision-language foundation models extend beyond segmentation. BiomedGPT supports diverse biomedical tasks, RadFM incorporates large-scale two- and three-dimensional medical data, CT-CLIP and CT-CHAT use paired volumetric CT and reports, and Merlin aligns CT volumes with reports and structured electronic health record data. Collectively, these studies demonstrate that paired clinical text can provide scalable supervision that reduces dependence on some manually curated labels, while supporting retrieval, classification, and generative interaction. They do not establish that a generic radiology model understands tracer physiology or quantitative PET (6–9). Clinically focused applications of imaging foundation models are also emerging outside PET/CT; for example, a CT-based deep foundation model has been evaluated for prediction of immune checkpoint inhibitor–induced pneumonitis risk in lung cancer (46), illustrating the use of a foundation-model approach for a narrowly defined oncology endpoint. Figure 2 presents a proposed PET/CT foundation-model architecture that aligns molecular imaging, anatomic CT, clinical text, and longitudinal data within a shared representation while retaining modular, traceable outputs for segmentation, retrieval, quantification, response assessment, and report generation.

A useful distinction is therefore needed. A large pretrained model is not automatically a clinically useful foundation model, and a model described as generalist may remain limited to one modality or anatomic region. Foundation status should be supported by evidence of task breadth, adaptation efficiency, cross-domain transfer, robust external performance and transparent limitations. Recent medical foundation-model studies similarly warn against equating parameter count with generalizability (14–17).

3.5 Operational taxonomy of PET/CT foundation-model systems

For consistency, we classify PET/CT AI systems using the operational taxonomy summarized in Table 2, based on demonstrated capabilities rather than model size or terminology. The taxonomy distinguishes task-specific or transfer models, reusable pretrained representations, promptable foundation models, multimodal or vision-language foundation models, generalist clinical foundation models, and agentic systems. Capability class is considered separately from evidence maturity and clinical readiness.

4 PET-native multimodal and vision-language models

4.1 Promptable and pretrained PET representations

PET-specific foundation models began to emerge in 2025. SegAnyPET introduced a three-dimensional promptable segmentation model trained on PETS-5k (5,731 whole-body FDG PET volumes). It achieved mean Dice similarity coefficients of 90.49%–91.05% for prompted organ segmentation in its evaluated setting, including organs seen during training; performance was lower for unseen organs and markedly lower for zero-shot multifocal tumor segmentation. It should therefore be interpreted as a promptable PET representation rather than evidence of universal lesion segmentation (47). A parallel cross-modal effort, FratMAE, paired separate PET and CT vision-transformer encoders with cross-attention masked-autoencoder pretraining, providing an early demonstration that anatomic and molecular streams can be fused within a shared pretraining objective rather than concatenated post hoc; as a preprint focused on a defined task family, it still requires broader external and cross-tracer evaluation (48). In 2026, an open multicenter FDG PET/CT foundation model used 4,797 scans from four public datasets and held out 200 AutoPET scans for testing. With a SwinUNETR-v2-base backbone, masked-autoencoder initialization improved Dice from 0.4889 to 0.6132 with 70 labeled training scans and from 0.6414 to 0.7205 with 700; independent non-FDG and prospective validation remain absent (49). Prompt-agent reconstruction and cross-scanner diffusion denoising reuse pretrained priors but are classified conservatively as task-specific or transfer systems within this taxonomy because broad cross-task adaptation has not been demonstrated (19, 30).

4.2 PET/CT vision-language grounding and report generation

One of the most rapidly emerging frontiers is PET/CT vision-language modeling. PET reports contain lesion locations, tracer avidity, SUV values, comparison with prior examinations and an impression that integrates disease distribution. A model must therefore connect words to three-dimensional, often multifocal findings. ConTEXTual Net 3D used weak labels derived from report mentions and scan coordinates to train text-guided PET/CT visual grounding. This approach converts routine reports into scalable spatial supervision, although report omissions and automatically generated labels can introduce systematic bias (50).

PET2Rep established a benchmark of 565 whole-body PET/CT–report pairs and evaluated 30 general-purpose and medical vision-language models. Its best reported overall reference score was 26.8%, compared with 27.5% for a fixed-template baseline; this composite score summarizes heterogeneous language and clinical-efficacy metrics and should not be interpreted as an accuracy percentage. Practically, the result indicates that the best evaluated model did not outperform the fixed-template baseline on the benchmark's aggregate measure, underscoring limitations in the performance of currently evaluated PET/CT report-generation models (51). PETAR aligned PET, CT, lesion masks, and 11,356 localized descriptions from 5,126 examinations. PETAR-4B reported GREEN 0.257 and RaTEScore 0.713, and five physicians preferred or tied its outputs in 69 of 116 blinded comparisons. However, numeric values were masked during training and quantitative errors were excluded from the reader evaluation, limiting conclusions about SUV or lesion-size fidelity (52). PETRG-3D similarly uses separate PET and CT branches and style-adaptive prompts for a multicenter lymphoma corpus comprising 824 PET/CT reports (53).

Moving from task benchmarks toward a self-described generalist system, the SDF-HOLO used 10,460 PET/CT–report studies from four tertiary hospitals in China and combined dual-stream anatomical–metabolic encoding with voxel–text alignment for segmentation, detection, and multilingual reporting. It reported AutoPET Dice values of 0.6328–0.6933 and ROUGE-L of 0.4608 (54). A parallel 2026 pilot study used a large language model to orchestrate validated tools for end-to-end PET/CT interpretation and quantification. This agentic pattern may offer a more modular and inspectable architecture than asking a single monolithic model to perform every operation internally: segmentation, SUV measurement, structured reasoning and drafting can remain versioned, testable and replaceable. The pilot remains early evidence and should not be interpreted as support for autonomous clinical use (55). Representative studies are summarized in Table 3, which distinguishes reported quantitative performance from evidence maturity. Although several systems demonstrate strong technical performance, none has yet undergone prospective clinical-impact evaluation.

Table 3. Representative studies marking the transition toward PET-native foundation and vision-language models.

Dataset sizes, publication status and evidence gaps reflect the cited reports as of July 11, 2026. Because these studies evaluate different tasks, datasets, endpoints, and performance metrics (e.g., Dice, F1, ROUGE-L, and GREEN), the reported values are study-specific and should not be interpreted as directly comparable measures of model performance.

4.3 Current evidence gaps

Four gaps are conspicuous. First, most PET foundation models focus on FDG oncology; evidence for neurodegenerative, cardiac, pediatric, infection/inflammation and theranostic applications is sparse. Second, longitudinal reasoning is underdeveloped even though response assessment depends on consistent comparison across time. A longitudinally aware network for pediatric Hodgkin lymphoma shows the task-specific potential of serial modeling, but generalist longitudinal PET benchmarks remain absent (56). Newer resources are beginning to close the report-linked gap; for example, PETWB-REP releases whole-body FDG PET/CT from 490 patients spanning multiple cancers with paired radiology reports, although it remains single-institution and FDG-only (57). Table 4 maps the principal datasets and benchmarks supporting this emerging ecosystem and highlights the limited tracer diversity, longitudinal linkage and report-level supervision available for foundation-model evaluation.

Table 4. Key datasets and benchmarks supporting the PET/CT foundation-model ecosystem, with their principal uses and current limitations.

Third, the relationship between text generation and quantitative correctness is unresolved (50–53). A report may correctly mention a lesion but misstate its SUV, size, temporal change or anatomic location. Fourth, prospective human-AI evidence is rare. The field has many retrospective accuracy studies but few silent deployments, crossover reader trials, time-motion studies or patient-outcome evaluations (13, 55, 56).

5 Clinical opportunities and translation challenges

The opportunities outlined below remain prospective and should not be interpreted as established clinical benefits; their translation into routine practice will require task-appropriate external validation and, where relevant, prospective clinical evaluation.

5.1 Acquisition and quantitative image formation

A pretrained molecular-imaging encoder could adapt to low-count denoising, reconstruction, motion correction and quality control across scanners and protocols. Total-body PET may provide a useful setting for representation learning because it produces high-sensitivity dynamic and whole-body data at a scale that could support such approaches. Potential clinical benefits could include lower activity, faster scans, and improved access. A key translational requirement is preservation of quantitative fidelity at lesion, organ and kinetic levels (2, 27, 28).

5.2 Universal segmentation and tumor-burden phenotyping

Promptable segmentation may enable automatic models to function as interactive physician tools (42–44). A user could specify a lesion, organ or physiologic region and obtain an editable three-dimensional contour, rather than accepting a fixed tumor mask. Shared representations could also support lesion tracking, total metabolic tumor volume, organ-specific disease burden, and radiopharmaceutical dosimetry. The interface should expose uncertainty and any gains in correction efficiency should be evaluated against manual contouring.

5.3 Reporting, retrieval and longitudinal comparison

Vision-language alignment may support retrieval of similar cases, grounding of report phrases to lesions, structured draft generation, and automated comparison with prior scans (5, 6, 50, 52). For response assessment, the system should pair each current lesion with its prior counterpart, calculate quantitative change, identify new lesions and generate an evidence-linked draft. Every generated statement should be traceable to an image region, quantitative tool output or clinical source.

5.4 Theranostics and precision oncology

Nuclear medicine uniquely connects diagnosis and therapy. A multimodal model could integrate target-expression PET, tumor burden, organ dosimetry, laboratory trends, prior treatments and outcomes to support patient selection and treatment planning. Such systems could help standardize complex workflows, but causal treatment recommendations require stronger evidence than diagnostic assistance. Retrospective associations should not be presented as evidence that a model can select the best radiopharmaceutical or activity.

5.5 Agentic orchestration as a near-term architecture

The near-term clinical system is likely to be a supervised orchestrator rather than an autonomous foundation model. A language model can route a case through DICOM validation, registration, organ and lesion segmentation, SUV extraction, response rules, retrieval and report drafting. Each tool can retain its own version, validation record and failure checks. The nuclear medicine physician remains responsible for reconciling discordant evidence and approving the report. Early PET/CT agent evidence supports feasibility but not autonomous deployment (55).

5.6 Risks and unresolved technical problems

5.6.1 Hallucination and semantic overconfidence

Medical vision-language models can generate convincing but unsupported details. In one evaluation of GPT-4V across radiologic images, false-positive rates were high and performance for the tasks studied remained well below expert readers. PET reporting adds further opportunities for hallucination through tracer name, SUV, lesion count, comparison and staging language. A clinically safe model should abstain, cite its evidence and separate measured facts from inferred impressions (13).

5.6.2 Quantitative drift hidden by visual quality

A foundation model may adapt well to classification while altering SUV distributions or lesion texture (11, 12, 36). Quantitative drift can result from scanner upgrades, reconstruction changes, dose protocols or model updates (10). PET validation should include Bland-Altman analysis, repeatability coefficients, lesion-size strata, high- and low-uptake lesions, physiologic reference organs and response-category concordance.

5.6.3 Cross-tracer negative transfer

Pooling tracers without explicit conditioning can create negative transfer (26, 27). A model may learn that intense salivary-gland uptake is suspicious from one training context and normal from another or conflate treatment-related and pathologic patterns. Tracer identity, injected activity, uptake interval and acquisition protocol should be explicit inputs (49). Evaluation should report performance by tracer and by tracer-disease combination, including rare combinations.

5.6.4 Dataset leakage and report-derived label bias

Foundation-model corpora create new leakage pathways (14–17). The same patient may appear across serial scans, public dataset derivatives or training and evaluation reports. Report-derived labels also reproduce the omissions, hedging, stylistic variation and prior assumptions of the reporting practice. Patient-level and institution-level deduplication, temporal splits, source provenance and contamination audits are essential.

5.6.5 Fairness, privacy and access

Scanner availability, tracer access and referral patterns vary across health systems. A model trained at tertiary cancer centers may underperform in community practice, in pediatric patients or in populations with different disease prevalence. Large multimodal data also increases re-identification risk because images, text and longitudinal metadata can be linked. Federated learning, secure computation and de-identification can reduce some risks, but governance must define permissible use, access, retention and accountability (38, 58).

6 MOSAIC-PET framework and translational roadmap

6.1 MOSAIC-PET clinical-readiness framework

We propose MOSAIC-PET as an author-developed, task-agnostic framework for evaluating PET/CT foundation models across reconstruction, segmentation, retrieval, reporting, prediction, and agentic workflows. It is intended to complement, rather than replace, CLAIM 2024 for imaging-AI reporting, TRIPOD + AI for prediction models, FUTURE-AI for trustworthy lifecycle principles, DECIDE-AI for early live clinical evaluation, REAL-FM for evaluating biomedical foundation models claims, and REFINE for foundation- and large-language-model reporting (58–63). MOSAIC-PET adds PET-specific requirements for quantitative molecular fidelity, image–text grounding, tracer-conditioned adaptation, clinical integration, and lifecycle control. Its six domains are a structured synthesis of the reviewed literature and the authors' judgment; they have not undergone Delphi consensus, weighting, threshold validation, or prospective testing. The framework must therefore not be used as a certification score. Figure 3 introduces the domains, and Table 5 translates them into intended-use-specific questions, minimum evaluations, and illustrative stop criteria.

Table 5. The proposed operationalization of MOSAIC-PET.

The domains, minimum evaluations and example stop criteria are intended as a proposed evaluation framework; thresholds should be prespecified according to the intended use and clinical risk.

6.2 Translational roadmap

Figure 4 summarizes five macro-stages from broad PET/CT pretraining to clinical impact: build the data foundation, define the intended use, test clinically meaningful performance, deploy progressively, and govern the total product lifecycle. Sections 6.2.1–6.2.6 provide six implementation actions mapped to those five stages; human-factors and equitable-implementation validation is separated as a cross-cutting action spanning clinical testing and progressive deployment. Accumulating evidence and continuous monitoring connect the stages, emphasizing that scale alone does not establish clinical readiness.

6.2.1 Build multi-tracer, longitudinal and report-linked corpora

The field requires datasets that span institutions, scanner vendors, reconstruction methods, disease states, and radiotracers. Each dataset should include de-identified acquisition metadata, injected activity, uptake time, reconstruction and correction parameters, quantitative reference regions, lesion annotations where available, clinical reports, and relevant outcomes. Longitudinal examinations should be linked at the patient level, with data partitioned to prevent patient-level and temporal leakage. Public challenges have demonstrated how shared benchmarks can accelerate segmentation research; the next generation should incorporate report-linked data, serial examinations, and multi-tracer diversity (32, 33, 35).

6.2.2 Separate pretraining breadth from intended-use evaluation

A foundation model may be pretrained on heterogeneous data, but its regulatory and clinical claims should remain specific to the intended use (58, 61, 63). Developers should clearly define the intended user, inputs, outputs, target population, radiotracer, workflow position, and supported clinical action. Adaptation and testing datasets must remain independent of pretraining data, and potential data contamination should be systematically assessed and reported. Evidence should also distinguish among zero-shot performance, parameter-efficient adaptation, and full model fine-tuning.

6.2.3 Evaluate clinically meaningful performance, safety, and utility

Dice coefficient, AUROC, BLEU and ROUGE provide useful technical information but are insufficient to establish clinical value. Lesion-segmentation studies should assess false-positive volume, lesion-level sensitivity and agreement in total tumor burden (35, 36, 47, 64). Prediction models require calibration, external testing and clinical-utility analysis (22, 37, 59). Report-generation systems should be evaluated for factual accuracy, completeness, anatomic localization, quantitative correctness, negation, uncertainty and longitudinal consistency (50–52, 55). Reconstruction, enhancement, and generative image-synthesis models require quantitative equivalence, lesion-level analysis, and reader-assessed noninferiority (10–12, 21). Evaluation datasets should also include challenging cases in which the model is expected to abstain or request expert review (58, 63).

6.2.4 Validate human factors, workflow integration, and equitable implementation

Clinical translation requires evidence that extends beyond standalone model performance. Human-factors studies should evaluate usability, interpretability, alert and correction burden, automation bias, appropriate reliance, and the ability of clinicians to recognize and override erroneous outputs. Evaluation should include nuclear medicine physicians with different experience levels and should examine whether performance varies across institutions, demographic groups, disease presentations, radiotracers, scanner platforms, and resource settings. Implementation studies should additionally assess interoperability with PACS, reporting systems, and clinical workflows; cybersecurity requirements; staff training; and cost-effectiveness. These considerations are particularly important for foundation models because their broad capabilities may create new forms of interaction, overreliance, and performance disparity that are not captured by conventional retrospective benchmarks (58–61).

Computational feasibility should also be considered in assessments of clinical readiness. Training, adapting, or locally deploying large multimodal foundation models may require substantial accelerator hardware and memory, storage capacity, network bandwidth, and technical support, which may limit adoption in smaller or resource-constrained nuclear medicine services. Centralized or cloud-based inference may reduce local hardware requirements but may introduce dependencies on connectivity, latency, data governance, and recurring infrastructure costs (65). Accordingly, translational evaluations should report hardware requirements, inference time, storage and memory demands, deployment architecture, and maintenance burden, and should consider parameter-efficient adaptation, model compression, or modular workflows when these can achieve the intended clinical function with lower computational requirements (5).

6.2.5 Progress from silent evaluation to assisted use and clinical-impact studies

A rational deployment pathway should begin with retrospective external testing, followed by prospective silent evaluation to characterize real-world performance, data drift, and failure modes without influencing clinical decisions (58, 66). The subsequent assisted phase should assess reader accuracy, interpretation time, correction burden, and automation bias through crossover or randomized studies. Only after these requirements are satisfied should studies investigate effects on clinical management, patient outcomes, and resource utilization. Performance monitoring must continue after deployment because scanners, radiotracers, acquisition protocols, and patient populations change over time.

6.2.6 Align reporting, trustworthiness and regulation

PET/CT foundation-model studies should use CLAIM 2024 for imaging-AI reporting, TRIPOD + AI when a prediction model is developed or evaluated, FUTURE-AI for trustworthy lifecycle principles, REFINE for foundation- and large-language-model reporting, and DECIDE-AI for early-stage prospective evaluation in live clinical settings (58–62). In the United States, the FDA's January 2025 draft guidance describes lifecycle-management and marketing-submission recommendations for AI-enabled device software functions; it remains draft guidance. The FDA's August 2025 final guidance on predetermined change control plans addresses prespecified modifications to AI-enabled device software functions (66, 67). Regulatory submissions and institutional governance should therefore address intended use, update boundaries, performance monitoring, cybersecurity, clinician- and patient-facing information, auditability, and rollback procedures rather than treating authorization as a one-time event. Figure 3, Table 5, and Figure 4 connect PET/CT-specific evidence domains with claim-specific testing, specialist-supervised deployment, and lifecycle control.

Patient-level transparency should accompany technical and regulatory governance. When AI meaningfully contributes to image interpretation, quantitative assessment, or clinical decision support, patients should receive information appropriate to the intended use regarding the role of AI, the continued responsibility of the clinical team, and relevant limitations or uncertainties. Requirements for specific consent or notification may vary according to the clinical context, jurisdiction, institutional policy, secondary use of patient data for model development or evaluation, and the degree of system autonomy. Institutions should therefore define how patient-facing information is communicated, how questions or concerns are addressed, and, where applicable, how consent or opt-out mechanisms are implemented. These considerations should be incorporated into implementation and governance planning alongside privacy, safety, performance monitoring, and accountability.

6.3 Limitations of this review

This review has several limitations. First, it was designed as a focused narrative review rather than a systematic review; therefore, the search strategy was not intended to achieve exhaustive study capture or formal risk-of-bias assessment. Second, the search was restricted to English-language literature identified through PubMed, arXiv, and official conference proceedings repositories, potentially excluding relevant studies indexed elsewhere or published in other languages. Third, given the rapidly evolving nature of PET/CT foundation-model research, selected recent preprints were included. These studies were explicitly identified as preprints and interpreted as preliminary evidence, as their findings may change following peer review. Fourth, general medical-imaging foundation-model studies were included selectively when they provided directly relevant methodological or translational context, which necessarily introduced an element of author judgment. Finally, the literature search covered studies available through July 11, 2026; subsequent developments are therefore not represented. Accordingly, this review should be interpreted as a structured synthesis of the contemporary evidence landscape rather than an exhaustive or formally graded systematic review.

7 Conclusions and future directions

The reviewed evidence supports a transition in development strategy, not yet a transition to routine clinical practice. Task-specific PET/CT systems have established technical feasibility for reconstruction, enhancement, segmentation, quantification, and outcome modeling. More recent promptable, pretrained, vision-language, and agentic systems broaden the range of reusable components, but the evidence base is heterogeneous and dominated by retrospective cohorts, task-specific endpoints, FDG oncology, and single-system datasets. Technical performance across Dice, F1, language metrics, reader ratings, and staging measures is not directly comparable and does not by itself demonstrate clinical utility. The most defensible near-term architecture is therefore modular and claim-specific: PET physics, quantitative tools, image–text grounding, and language generation should remain separately testable, evidence linked, and overseen by nuclear medicine specialists. This is a reasoned translational proposal, not a prediction validated by comparative deployment trials. Future work should prioritize multitracer and longitudinal PET/CT corpora with acquisition metadata and institutionally sequestered test sets; PET-specific factuality benchmarks for tracer identity, localization, lesion count, SUV, size, negation, uncertainty, and temporal change; and prospective studies that progress from silent evaluation to reader-assistance and clinical-impact endpoints. Dynamic and total-body PET, theranostic dosimetry, human factors, interoperability, compute burden, and post-deployment drift require explicit evaluation. A credible PET/CT foundation-model ecosystem should preserve quantitative molecular meaning, expose uncertainty, ground language in images and measurements, generalize across clinically relevant domains, and remain auditable throughout its lifecycle. MOSAIC-PET offers a proposed structure for organizing that evidence, but its own clinical value must be established empirically.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. ChatGPT (OpenAI; GPT-5.6 Sol) was used to assist with language refinement, grammar, clarity, and phrasing during manuscript preparation. All AI-assisted content was critically reviewed, edited, and verified by the authors, who take full responsibility for the accuracy, originality, scientific integrity, and final content of the manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Vincent Habouzit, Centre Hospitalier Universitaire (CHU) de Saint-Étienne, France

Reviewed by: Batyrkhan Omarov, Al-Farabi Kazakh National University, Kazakhstan

Dawa Chyophel Lepcha, Taipei Medical University, Taiwan