Artificial intelligence in computed tomography imaging for pulmonary embolism: a narrative review from computed tomography pulmonary angiography to non-contrast computed tomography

Introduction

Pulmonary embolism (PE) is a serious clinical condition associated with substantial morbidity and mortality and is often considered the third most common major cardiovascular disease after coronary heart disease and stroke (1-4). Its clinical manifestations are diverse and often nonspecific, including chest pain, dyspnea, hemoptysis, syncope, or even asymptomatic presentations in selected patients (1,5). Because delayed diagnosis and treatment may lead to adverse outcomes, timely and accurate imaging assessment remains central to PE management.

Computed tomography pulmonary angiography (CTPA) is currently the diagnostic reference standard for PE and is central to diagnostic pathways, with high reported sensitivity and specificity (6-8). However, CTPA imaging can be affected by inadequate pulmonary arterial opacification and patient-specific issues related to iodinated contrast use, including renal insufficiency or allergy (1,9). In contrast, non-contrast computed tomography (NCCT) avoids iodinated contrast administration and may offer value in selected patients who cannot undergo CTPA. Nevertheless, unenhanced computed tomography (CT) has lower diagnostic performance for PE than CTPA, particularly for small or peripheral thrombi, because intravascular contrast is absent and thrombus conspicuity is limited (10).

Artificial intelligence (AI), particularly deep learning, has introduced new opportunities for PE assessment on CT imaging. In CTPA, AI models can assist thrombus detection, localization, segmentation, clot burden quantification, workflow triage, and risk stratification (11-15). The high vascular contrast of CTPA provides a favorable imaging substrate for precise thrombus localization and quantitative analysis (16-21). By contrast, NCCT presents a substantially more challenging environment for AI because of lower vascular contrast and reduced thrombus visibility. These modality differences motivate indirect sign analysis, synthetic contrast generation, and transmodal or cross-modality learning strategies, which are reviewed in detail below. The overall AI-assisted diagnostic workflow across CTPA and NCCT is summarized in Figure 1.

Figure 1 AI-assisted PE diagnosis pipeline across CTPA and NCCT modalities. The diagram illustrates the flow from CT image acquisition through AI clinical tasks and cross-modality strategies to clinical outputs, with cross-cutting challenges annotated as a surrounding framework rather than a pipeline endpoint. AI, artificial intelligence; CT, computed tomography; CTPA, computed tomography pulmonary angiography; NCCT, non-contrast computed tomography; PE, pulmonary embolism; RV/LV, right ventricle-to-left ventricle.

This narrative review addresses three key questions: (I) what AI approaches have been developed and evaluated for CTPA-based PE diagnosis, and what clinical tasks do they currently support? (II) What strategies are emerging to enable AI-based PE assessment on NCCT, and how do they compare in evidence maturity? (III) What translational barriers and evidence gaps must be addressed to move from algorithmic innovation to routine clinical implementation? We present this article in accordance with the Narrative Review reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0293/rc).

Methods

A structured literature search was performed in accordance with narrative review best practices. The initial search was performed on June 30, 2025 and updated on December 4, 2025. An additional search update was conducted in June 2026 during revision to identify recently published studies directly relevant to NCCT-based PE assessment, synthetic contrast generation, and cross-modality learning. Eligible publications from database inception to June 2026 were considered. Databases queried included PubMed, Google Scholar, Web of Science, and IEEE (Institute of Electrical and Electronics Engineers) Xplore. PubMed was searched using a combination of Medical Subject Headings (MeSH) and free-text terms related to PE, CT pulmonary angiography, non-contrast CT, AI, deep learning, machine learning, computer-aided diagnosis, and radiomics. The complete PubMed search strategy is provided in Table S1, and detailed search parameters are summarized in Table 1.

Inclusion criteria were: (I) peer-reviewed original research articles, validation studies, or peer-reviewed conference proceedings; (II) focus on AI and deep learning applications for PE detection, diagnosis, segmentation, quantification, triage, or risk stratification using CTPA or NCCT; (III) reporting of quantitative findings or clinically relevant methodological insights; and (IV) English-language publications. Studies on NCCT-to-contrast-enhanced CT synthesis or cross-modality learning were also considered when directly relevant to PE assessment or to the diagnostic safety of synthetic imaging. Selected review articles or editorials were used only for contextual discussion of clinical background, reporting standards, or translational challenges, and were not treated as primary diagnostic performance evidence. Exclusion criteria were: (I) non-imaging AI applications; (II) studies without modality-specific analysis relevant to CT-based PE assessment; and (III) non-peer-reviewed content, unless clearly identified as emerging evidence and used only for contextual discussion. Screening followed a two-stage process: initial title and abstract review by two independent reviewers (L.L. and C.L.), followed by full-text assessment. Disagreements were resolved by discussion and consensus with a senior reviewer (M.Y.). Reference lists of included studies were hand-searched for additional relevant literature. Data extraction captured study design, sample size, imaging modality, AI architecture or strategy, performance metrics, validation design, imaging protocols, and reported limitations.

AI applications in CTPA-based PE diagnosis

CTPA provides the best-established imaging substrate for AI-assisted PE analysis because emboli can be directly visualized within contrast-opacified pulmonary arteries. Accordingly, CTPA-based AI has evolved from candidate detection toward triage, segmentation, clot burden quantification, and risk stratification. In this section, we organize AI applications into three clinical-task-driven categories: detection and triage, segmentation and quantification, and risk stratification and decision support. Detailed performance metrics for individual studies are summarized in Table 2.

Detection and workflow triage

Early computer-aided detection systems demonstrated the feasibility of automated PE detection using hand-crafted three-dimensional vascular and geometric features (22,23). However, these approaches were constrained by reliance on manually engineered features, high false-positive rates, and computational complexity—limiting their clinical applicability (Table 2).

The transition to deep learning enabled end-to-end volumetric analysis and substantially greater automation. Convolutional neural networks, object-detection frameworks, and vessel-oriented image representations have been evaluated across multiple architectures and datasets (24-26,28-30). Across these heterogeneous studies, two consistent patterns emerge: detection performance is generally higher for central and lobar pulmonary emboli, whereas subsegmental PE and artifact-related false positives remain persistent challenges irrespective of the specific architecture employed. Detailed study-level metrics are provided in Table 2.

More recent AI systems have moved beyond binary detection toward clinical workflow integration. AI-based worklist triage and reprioritization have demonstrated substantial reductions in time to diagnosis in real-world deployments (31-33), reflecting a shift from algorithm development toward measurable clinical impact (Table 2).

Taken together, these studies indicate that CTPA-based AI has moved beyond binary detection toward quantitative and workflow-oriented applications. However, the lack of standardized evaluation protocols across studies—including inconsistent definitions of false-positive tolerance, variable PE location stratification, and heterogeneous validation designs—continues to limit direct cross-study comparison (Table 2).

Segmentation and clot burden quantification

U-Net-based architectures have become the predominant approach for PE segmentation, enabling not only thrombus localization but also automated clot burden quantification (34-36). Compared with early candidate detection systems, segmentation models provide voxel- or lesion-level outputs that can be linked to established visual scoring systems, such as the Qanadli and Mastora scores. For example, deep learning-derived clot burden has been shown to correlate strongly with conventional manual scores and right ventricular dysfunction parameters, supporting its potential value for objective and reproducible severity assessment (36). Large multi-institutional datasets, such as RSPECT, further indicate that nnU-Net-based approaches can scale from lesion detection to clot volume quantification and severity correlation (18).

Recent model development has extended beyond conventional U-Net designs. Attention-enhanced and hybrid architectures, including Swin Transformer- and convolutional neural network-based models, are designed to combine global contextual modeling with local spatial feature extraction, whereas artery-aware and multi-task learning frameworks incorporate vascular context to support simultaneous detection, localization, acute/chronic classification, and right ventricle-to-left ventricle (RV/LV) ratio estimation (37,38,42). These approaches reflect a broader shift from isolated thrombus segmentation toward integrated anatomical and clinical characterization.

Weakly supervised and semi-weakly supervised learning approaches are particularly relevant because pixel-level PE annotation is labor-intensive and subject to inter-reader variability. Studies using weak labels or reduced annotation requirements have shown that competitive diagnostic performance can be achieved without exhaustive manual thrombus delineation (15,39,43).

Nevertheless, segmentation performance remains moderate across studies, and subsegmental or peripheral emboli continue to be more difficult to segment reliably than thrombi in the main or lobar pulmonary arteries (Table 2).

Risk stratification and clinical decision support

Beyond detection and segmentation, AI models incorporating clinical variables and imaging-derived biomarkers have shown promise for PE assessment and risk stratification. Models integrating CTPA features with clinical data or electronic medical record variables have reported improved discriminatory performance compared with single-modality approaches in retrospective settings (40,41). Similarly, multi-task models that combine PE detection with localization, acute/chronic classification, and RV/LV ratio estimation may support more comprehensive clinical decision support than binary PE classification alone (42).

Several studies have further explored AI-derived thrombus burden and cardiac parameters as quantitative markers for PE risk stratification. Deep learning-derived clot burden metrics, clot ratio, and automated Qanadli- or Mastora-like measurements have been reported to correlate with clinical risk strata and visual scoring systems (19-21). These quantitative tools may improve reproducibility compared with subjective visual estimation and may help identify patients requiring closer monitoring or treatment escalation. However, prospective validation demonstrating clinical utility and patient-level benefit remains limited.

Summary of CTPA evidence and remaining gaps

Overall, CTPA-based AI has accumulated a relatively mature evidence base compared with NCCT-based AI. AI models have achieved high reported sensitivity for central and lobar PE across multiple architectures and selected multicenter or external validation settings, and AI-based workflow triage has reduced time to diagnosis in real-world deployments. Automated clot burden quantification also shows meaningful correlations with established manual scoring systems and right ventricular dysfunction parameters. Nevertheless, important gaps remain: subsegmental PE detection and segmentation remain challenging; many studies are retrospective or single-center; prospective reader and workflow studies are still limited; and standardized benchmark datasets with vessel-level PE annotations are lacking. These issues are examined further in the Translational Challenges section.

AI applications in NCCT-based PE assessment

NCCT may serve as a contrast-free imaging option for selected patients who cannot undergo iodinated contrast-enhanced CTPA. However, its inherently low vascular contrast poses substantial challenges for PE diagnosis. AI strategies for NCCT-based PE assessment have therefore evolved along several complementary pathways, each addressing the low-contrast problem differently. We organize these strategies into indirect sign analysis, direct AI detection, synthetic contrast-enhanced CT generation, and cross-modality knowledge transfer. Key studies are summarized in Table 3.

Indirect signs and quantitative NCCT markers

Although NCCT lacks the direct thrombus visualization afforded by CTPA, several indirect imaging signs and quantitative measurements may provide diagnostic clues. Early reports described hyperdense lumen signs in central PE, and subsequent studies evaluated thrombus attenuation, thrombus-to-blood-pool ratio, pulmonary artery dilation, and peripheral wedge-shaped opacities as potential NCCT markers (44-47). These studies collectively suggest that quantitative attenuation-based measurements and systematic assessment of multiple indirect signs may improve PE probability estimation, particularly for central emboli.

However, the diagnostic value of NCCT remains limited by low sensitivity for peripheral PE. Across available studies, specificity for central PE-related signs is generally higher than sensitivity, whereas peripheral or subsegmental emboli remain difficult to identify reliably on native NCCT. Therefore, indirect signs and quantitative NCCT markers may support triage or risk stratification in selected patients, but NCCT alone cannot reliably exclude PE or replace CTPA when contrast-enhanced imaging is feasible (Table 3).

Direct AI detection on NCCT

Direct application of deep learning to NCCT for PE detection represents a conceptually straightforward but technically challenging strategy. Hagen et al. (48) developed a deep learning-based AI model for PE detection on NCCT and evaluated its performance in 99 cancer patients. The algorithm achieved sensitivities of 54.5%, 81.9%, and 80.0% for central, segmental, and subsegmental PE, respectively—an intriguing pattern in which sensitivity was higher for smaller, more peripheral emboli than for central ones. This counterintuitive distribution may reflect differences in local tissue contrast: central pulmonary arteries are surrounded by mediastinal soft tissue, which reduces thrombus conspicuity on NCCT, whereas segmental and subsegmental arteries are surrounded by aerated lung parenchyma providing greater natural contrast (48). However, this observation should be interpreted cautiously because current NCCT datasets remain small (Hagen et al. included only 99 cases from a single center), and whether this pattern represents a reproducible modality-specific phenomenon or a dataset-dependent observation remains uncertain.

Synthetic contrast-enhanced CT generation

A growing body of work has explored deep generative models, predominantly generative adversarial network-based approaches, to synthesize contrast-enhanced CT from NCCT as a preprocessing step for PE assessment. This strategy aims to computationally restore vascular contrast that is absent in native NCCT images, thereby enabling downstream interpretation or AI models to operate on CTPA-like images without actual contrast administration.

Technical development has progressed from early NCCT-to-contrast synthesis models to more advanced multi-stage training strategies, transformer-enhanced generative adversarial network frameworks, transmodal learning strategies, and diffusion-based approaches (49-52). Across these studies, three key observations emerge. First, the training-data strategy may be as important as network architecture, as self-supervised pretraining, dual-energy CT pretraining, and registration-guided fine-tuning can each contribute to image synthesis performance. Second, global image quality metrics appear linked to downstream detection performance, suggesting that synthesis quality is not merely a visual endpoint but may affect diagnostic utility. Third, synthesis fidelity may vary by vessel size, raising concern that performance improvements may not be uniform across the pulmonary arterial tree.

PE-specific studies suggest that synthetic contrast generation may support PE detection, but they also highlight critical safety caveats. Perceptual realism alone does not guarantee lesion-level diagnostic fidelity, and diffusion-based models, although technically promising, still require validation against generative adversarial network-based approaches for PE-specific diagnostic outcomes. Detailed performance metrics are provided in Table 3, and diagnostic safety implications are discussed in Section “Diagnostic safety in synthetic imaging”.

Knowledge transfer and cross-modality learning

A related strategy involves transmodal or cross-modality learning, which aims to improve NCCT-based PE assessment by leveraging information across imaging domains. Recent transmodal frameworks suggest that such approaches may support automated PE diagnosis on NCCT, but the underlying mechanisms, data requirements, and generalizability require independent validation before clinical conclusions can be drawn (51).

Summary of NCCT evidence and safety concerns

NCCT-based AI for PE assessment remains at an earlier stage of evidence development than CTPA-based AI. Three complementary strategies have emerged: direct detection, synthetic contrast generation, and knowledge transfer. However, they have not been compared on a common dataset, and prospective validation remains absent. Synthetic contrast generation appears to have advanced rapidly, but cross-domain evidence suggests that visually convincing synthetic images may introduce imperceptible spurious features that degrade downstream task performance (53). These safety concerns underscore the need for lesion-level validation beyond global image similarity metrics. Although some recent NCCT AI studies include multicenter datasets, independent replication remains necessary.

Comparative roles of CTPA and NCCT in AI-assisted PE pathways

Imaging and clinical differences between modalities

CTPA and NCCT differ fundamentally in their imaging physics, clinical roles, and suitability for AI analysis. CTPA provides high-contrast opacification of the pulmonary arterial tree, enabling direct thrombus visualization across much of the pulmonary arterial tree and supporting AI-based segmentation, quantification, and risk stratification (6,54-58). However, CTPA requires intravenous iodinated contrast administration, which may be limited by allergic reactions, contrast-related risks in patients with severely reduced renal function, and other patient-specific contraindications (59-60). In contrast, NCCT avoids iodinated contrast administration and can be rapidly acquired, making it potentially useful as an adjunctive assessment option in selected patients who cannot undergo CTPA (10,32,61). The trade-off is substantially lower vascular contrast, which reduces direct thrombus conspicuity and makes PE assessment more dependent on indirect signs, quantitative attenuation measurements, or computational enhancement strategies (45,46,48). From an AI development perspective, CTPA provides a richer signal-to-noise environment for segmentation and quantification, whereas NCCT requires additional strategies—synthesis, knowledge transfer, or indirect feature analysis—to compensate for inherently low vascular contrast.

AI strategies across modalities: a conceptual framework

The AI strategies applicable to each modality differ in both technical approach and evidence maturity. For CTPA, AI applications have progressed along a relatively mature pathway, particularly for detection and workflow triage, with increasing extension toward segmentation, clot burden quantification, risk stratification, and clinical decision support. Multiple architectures have been evaluated across institutional settings, and some commercial AI tools for CTPA-based detection or triage have received regulatory clearance for clinical use. For NCCT, AI strategies can be conceptualized along four emerging but less mature pathways: (I) indirect sign analysis and quantitative marker measurement, which augment human interpretation of established NCCT features; (II) direct AI detection, which applies deep learning directly to native NCCT images; (III) synthetic contrast generation, which computationally reconstructs contrast-enhanced or CTPA-like vascular enhancement as a preprocessing step; and (IV) cross-modality knowledge transfer, which leverages paired CTPA-NCCT data during training to enable NCCT-only inference. These NCCT strategies differ substantially in their requirements for training data, computational resources, and regulatory pathways, and their relative clinical value remains to be established through comparative validation. A conceptual framework illustrating these pathways is provided in Figure 2, and the relative evidence maturity of the major CTPA- and NCCT-based strategies is summarized in Figure 3.

Figure 2 Conceptual framework for AI strategies across CTPA and NCCT modalities. Solid lines indicate relatively mature pathways; dashed lines indicate emerging pathways; dotted lines indicate early-stage pathways requiring independent validation. All NCCT AI strategies are presented as potential roles, not clinical recommendations. Positioning reflects evidence maturity and does not imply clinical superiority. AI, artificial intelligence; CT, computed tomography; CTPA, computed tomography pulmonary angiography; NCCT, non-contrast computed tomography; PE, pulmonary embolism.

Figure 3 Evidence maturity matrix for AI strategies in PE diagnosis. The matrix categorizes evidence by study design rather than diagnostic performance: √√√, substantial published evidence; √√, moderate published evidence; √, limited published evidence; limited, preliminary evidence from single studies; rare, isolated reports (1–2 studies); absent, no published evidence; emerging, recent findings not yet independently validated. Overall, CTPA-based strategies demonstrate greater evidence maturity than NCCT-based strategies, particularly for detection and workflow triage. AI, artificial intelligence; CTPA, computed tomography pulmonary angiography; NCCT, non-contrast computed tomography; PE, pulmonary embolism.

Evidence gaps in cross-modality comparison

Several evidence gaps are specific to the comparison between CTPA-based and NCCT-based AI approaches. First, no study has performed a head-to-head comparison of CTPA AI vs. NCCT AI for PE detection on the same patient cohort. Second, the lesion-level non-inferiority of synthetic contrast-enhanced CT or CTPA-like images compared with real CTPA for PE detection has not been established in a prospective trial. Third, the three emerging NCCT AI strategies—direct detection, synthetic contrast generation, and knowledge transfer—have not been compared on a common benchmark dataset with standardized PE annotations. Fourth, no cross-modality reader study has evaluated whether radiologist performance with NCCT plus AI assistance approaches that achieved with CTPA. These gaps represent important priorities for future research because they affect how future studies should define the clinical role of each imaging and AI strategy in selected patient populations.

Translational challenges

Despite significant progress, several interconnected challenges limit the clinical translation of AI for PE diagnosis. We group these challenges into six dimensions below.

Dataset scale and external validity

Many AI studies for PE detection—particularly NCCT-based studies—remain retrospective and single-center, with modest sample sizes. Although several larger CTPA datasets and real-world triage studies have emerged, external validation across institutions, scanner types, acquisition protocols, and clinical populations remains limited (62,63). Without multi-center validation using standardized evaluation protocols, the real-world performance of AI models for PE detection cannot be reliably estimated.

Annotation variability and reference standards

PE annotation is inherently subjective, particularly for small subsegmental emboli where inter-radiologist agreement is limited (64). Most studies use a single radiologist or consensus reading as the reference standard, which may introduce systematic bias. The lack of a standardized, publicly available annotation protocol—including definitions of positive PE at the vessel level, handling of partial filling defects, and minimum thrombus size thresholds—makes cross-study performance comparisons unreliable. This issue is compounded in NCCT studies, where thrombus boundaries are inherently ambiguous due to low contrast.

False positives and subsegmental PE detection

False positives in CTPA AI arise predominantly from contrast flow artifacts, pulmonary vein opacification, lymph nodes, and perivascular soft tissue—each mimicking thrombus appearance under certain imaging conditions (40). False negatives are concentrated in subsegmental and very small peripheral emboli, where partial volume effects and limited spatial resolution reduce detectability. Across multiple architectures and datasets, subsegmental PE detection sensitivity consistently lags behind that for central and lobar PE (27,44). This pattern has important clinical implications, as the therapeutic significance of isolated subsegmental PE remains debated, yet AI tools that systematically miss these findings may provide inappropriate reassurance.

NCCT-specific limitations

NCCT-based AI faces modality-specific challenges beyond those affecting CTPA. Low intrinsic contrast between thrombus and blood pool reduces the signal available for both human interpretation and AI analysis. Motion artifacts and partial volume effects are more consequential when the target-to-background contrast is already minimal. Some preliminary studies suggest that NCCT AI detection sensitivity may be lower for central PE than for peripheral PE (48)—potentially reflecting the loss of contrast between thrombus and surrounding mediastinal soft tissue—although whether this is a reproducible modality-specific pattern or a dataset-dependent observation requires validation in larger cohorts. NCCT is also less informative for severity assessment than CTPA because contrast-dependent signs, such as reflux into the inferior vena cava, cannot be evaluated, although some morphologic markers such as RV/LV ratio may still be measurable.

Diagnostic safety in synthetic imaging

For the synthetic contrast generation strategy, diagnostic safety represents one of the most critical translational concerns. Evidence from PE-specific and cross-domain studies highlights several risks. PulmoNet (51) suggested that global image quality metrics and downstream detection performance are closely associated, with lower structural similarity accompanied by reduced detection performance. In a cross-domain study of NCCT-to-contrast-enhanced CT synthesis for coronavirus disease 2019 (COVID-19) pneumonia assessment, Kalantar et al. (53) showed that visually convincing generative adversarial network-generated images could nevertheless degrade downstream radiomics and deep learning classification performance, indicating that synthetic images may contain imperceptible spurious features. For PE specifically, the most concerning potential failure modes include generation of spurious intravascular filling defects mimicking thrombi, obscuration of true thrombi by synthetic contrast, and vessel-caliber-dependent synthesis fidelity affecting detection performance non-uniformly across the pulmonary arterial tree. These risks underscore that clinical deployment of synthetic medical images should require uncertainty quantification, external validation, and lesion-level diagnostic safety assessment.

Interpretability, workflow integration, and regulatory issues

Many current AI models for PE detection, particularly deep learning architectures, provide limited uncertainty quantification and limited clinically interpretable explanations for their predictions. Few NCCT-based PE models report calibrated confidence estimates or mechanisms for flagging alternative diagnoses, which limits clinician trust and complicates medicolegal accountability. Although some CTPA-based triage tools have been integrated into clinical workflows, many research models remain offline and have not been tested in routine emergency radiology environments. From a regulatory perspective, while several CTPA-based AI tools for detection or triage have received U.S. Food and Drug Administration clearance, NCCT-based PE detection—particularly synthetic contrast generation—raises novel regulatory questions about the validation requirements for software that computationally modifies medical images before diagnostic interpretation. Prospective studies demonstrating improvement in patient-level outcomes—rather than diagnostic accuracy or turnaround time alone—remain lacking across PE AI applications.

Future directions

Based on the evidence synthesis and identified translational gaps, several priority areas for future research emerge. These priorities are organized by domain rather than timeline and correspond to four broad themes: data, model development, evidence generation, and clinical translation (Figure 4).

Figure 4 Research framework for AI-assisted PE diagnosis. Four parallel pillars—Data, Model, Evidence, and Translation—representing key domains for future investigation, not ranked by priority. AI, artificial intelligence; CTPA, computed tomography pulmonary angiography; NCCT, non-contrast computed tomography; PACS, picture archiving and communication system; PE, pulmonary embolism; RIS, radiology information system.

Data and benchmark development

One of the most important near-term priorities is the creation of publicly available, multi-center datasets with paired NCCT-CTPA examinations and vessel-level PE annotations. Such datasets are essential for comparing direct NCCT detection, synthetic contrast generation, and knowledge transfer approaches under a common reference standard. Standardized evaluation protocols should include stratification by PE location (central, segmental, subsegmental) and vessel caliber, as well as consistent definitions of false-positive tolerance and clinically acceptable detection thresholds.

Model development priorities

Improving subsegmental and peripheral PE detection remains a cross-architecture challenge that may benefit from dedicated high-resolution feature extraction strategies, synthetic data augmentation, or multi-scale attention mechanisms targeting small-vessel territories. For synthetic contrast generation, future studies should prioritize lesion-level and vessel-level validation over global image quality metrics alone. Uncertainty quantification should also be integrated into NCCT AI pipelines, particularly for synthetic imaging approaches, to help clinicians judge the reliability of AI-generated findings. Weakly supervised and semi-supervised methods that reduce annotation requirements merit further investigation, as evidence from CTPA studies suggests that a fraction of pixel-level annotations may be sufficient to approach fully supervised performance (43).

Evidence generation and clinical validation

Multi-center external validation across diverse CT scanner types, acquisition protocols, and patient populations should precede broad clinical implementation of NCCT-based AI strategies. Prospective reader studies comparing radiologist performance with and without AI assistance—for both CTPA and NCCT—would provide important evidence for clinical utility. Workflow integration trials should evaluate whether AI triage systems maintain their reported time-to-diagnosis reductions in real-world emergency department settings with heterogeneous case mix. For synthetic contrast generation, lesion-level non-inferiority studies comparing synthetic contrast-enhanced CT or CTPA-like images with real CTPA for PE detection are likely to be important for regulatory evaluation and clinical acceptance.

Translation and implementation

Federated learning frameworks may address data privacy concerns while enabling multi-institutional model development. Standardization of CT acquisition and reconstruction protocols for PE imaging—particularly for NCCT, where protocol variability substantially affects image quality—would facilitate cross-study comparability. Regulatory frameworks for AI-based image synthesis in diagnostic radiology, including validation requirements and post-market surveillance mechanisms, should be developed in parallel with technical advances. Interpretability methods that provide clinically meaningful explanations, such as attention maps overlaid on vascular anatomy and confidence scores stratified by embolus location, are needed to build clinician trust. Beyond acute PE, AI applications for chronic thromboembolic pulmonary hypertension remain underexplored and may represent a related area for future investigation (65).

Conclusions

This narrative review evaluated the evolving role of AI in PE diagnosis across CTPA and NCCT modalities. CTPA-based AI has accumulated relatively mature evidence for central and lobar PE detection, clot burden assessment, workflow prioritization, and risk stratification, although prospective multicenter studies demonstrating patient-level benefit remain limited. In contrast, NCCT-based AI remains at an earlier evidence stage. Three complementary strategies—direct detection, synthetic contrast generation, and cross-modality knowledge transfer—have emerged, each offering distinct opportunities but requiring further validation before clinical implementation.

Based on the current evidence, several priorities emerge: (I) consolidating CTPA-based AI through prospective multicenter workflow trials that move beyond single-center technical validation; (II) building shared NCCT-CTPA datasets with vessel-level PE annotations to enable systematic comparison of competing NCCT AI strategies; and (III) establishing lesion-level validation standards for synthetic imaging before clinical use. AI should be viewed as an augmentative tool that complements radiologist expertise through quantitative support and workflow prioritization rather than as a replacement for clinical interpretation. With rigorous validation, transparent uncertainty quantification, and human-centered implementation, AI may help optimize PE diagnostic pathways, reduce time to diagnosis, and extend diagnostic support to selected patients who cannot undergo contrast-enhanced CTPA, with the goal of improving patient outcomes.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the Narrative Review reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0293/rc

Funding: This work was supported by the National Natural Science Foundation of China (No. 82327803).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0293/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Konstantinides SV, Meyer G, Becattini C, Bueno H, Geersing GJ, Harjola VP, et al. 2019 ESC Guidelines for the diagnosis and management of acute pulmonary embolism developed in collaboration with the European Respiratory Society (ERS). Eur Heart J 2020;41:543-603. [Crossref] [PubMed]
2. Konstantinides SV, Barco S, Lankeit M, Meyer G. Management of Pulmonary Embolism: An Update. J Am Coll Cardiol 2016;67:976-90. [Crossref] [PubMed]
3. Payne JG, Tagalakis V, Wu C, Lazo-Langner A. CanVECTOR Network. Current estimates of the incidence of acute venous thromboembolic disease in Canada: A meta-analysis. Thromb Res 2021;197:8-12. [Crossref] [PubMed]
4. Calder KK, Herbert M, Henderson SO. The mortality of untreated pulmonary embolism in emergency department patients. Ann Emerg Med 2005;45:302-10. [Crossref] [PubMed]
5. Kearon C, de Wit K, Parpia S, Schulman S, Afilalo M, Hirsch A, Spencer FA, Sharma S, D'Aragon F, Deshaies JF, Le Gal G, Lazo-Langner A, Wu C, Rudd-Scott L, Bates SM, Julian JAPEGeD Study Investigators. Diagnosis of Pulmonary Embolism with d-Dimer Adjusted to Clinical Probability. N Engl J Med 2019;381:2125-34. [Crossref] [PubMed]
6. Stein PD, Fowler SE, Goodman LR, Gottschalk A, Hales CA, Hull RD, Leeper KV Jr, Popovich J Jr, Quinn DA, Sos TA, Sostman HD, Tapson VF, Wakefield TW, Weg JG, Woodard PK. PIOPED II Investigators. Multidetector computed tomography for acute pulmonary embolism. N Engl J Med 2006;354:2317-27. [Crossref] [PubMed]
7. Righini M, Le Gal G, Aujesky D, Roy PM, Sanchez O, Verschuren F, Rutschmann O, Nonent M, Cornuz J, Thys F, Le Manach CP, Revel MP, Poletti PA, Meyer G, Mottier D, Perneger T, Bounameaux H, Perrier A. Diagnosis of pulmonary embolism by multidetector CT alone or combined with venous ultrasonography of the leg: a randomised non-inferiority trial. Lancet 2008;371:1343-52. [Crossref] [PubMed]
8. Le Gal G, Righini M, Roy PM, Sanchez O, Aujesky D, Bounameaux H, Perrier A. Prediction of pulmonary embolism in the emergency department: the revised Geneva score. Ann Intern Med 2006;144:165-71. [Crossref] [PubMed]
9. Bae K, Jeon KN, Cho SB, Park SE, Moon JI, Baek HJ, Choi BH. Improved Opacification of a Suboptimally Enhanced Pulmonary Artery in Chest CT: Experience Using a Dual-Layer Detector Spectral CT. AJR Am J Roentgenol 2018;210:734-41. [Crossref] [PubMed]
10. Ehsanbakhsh A, Hatami F, Valizadeh N, Khorashadizadeh N, Norouzirad F. Evaluating the Performance of Unenhanced Computed Tomography in the Diagnosis of Pulmonary Embolism. J Tehran Heart Cent 2021;16:156-61. [Crossref] [PubMed]
11. Soffer S, Klang E, Shimon O, Barash Y, Cahan N, Greenspana H, Konen E. Deep learning for pulmonary embolism detection on computed tomography pulmonary angiogram: a systematic review and meta-analysis. Sci Rep 2021;11:15814. [Crossref] [PubMed]
12. Khan M, Shah PM, Khan IA, Islam SU, Ahmad Z, Khan F, Lee Y. IoMT-Enabled Computer-Aided Diagnosis of Pulmonary Embolism from Computed Tomography Scans Using Deep Learning. Sensors (Basel) 2023;23:1471. [Crossref] [PubMed]
13. Topff L, Ranschaert ER, Bartels-Rutten A, Negoita A, Menezes R, Beets-Tan RGH, Visser JJ. Artificial Intelligence Tool for Detection and Worklist Prioritization Reduces Time to Diagnosis of Incidental Pulmonary Embolism at CT. Radiol Cardiothorac Imaging 2023;5:e220163. [Crossref] [PubMed]
14. Ayobi A, Chang PD, Chow DS, Weinberg BD, Tassy M, Franciosini A, Scudeler M, Quenet S, Avare C, Chaibi Y. Performance and clinical utility of an artificial intelligence-enabled tool for pulmonary embolism detection. Clin Imaging 2024;113:110245. [Crossref] [PubMed]
15. Pu J, Gezer NS, Ren S, Alpaydin AO, Avci ER, Risbano MG, Rivera-Lebron B, Chan SY, Leader JK. Automated detection and segmentation of pulmonary embolisms on computed tomography pulmonary angiography (CTPA) using deep learning but without manual outlining. Med Image Anal 2023;89:102882. [Crossref] [PubMed]
16. Gotta J, Gruenewald LD, Martin SS, Booz C, Mahmoudi S, Eichler K, Gruber-Rouh T, Biciusca T, Reschke P, Juergens LJ, Onay M, Herrmann E, Scholtz JE, Sommer CM, Vogl TJ, Koch V. From pixels to prognosis: Imaging biomarkers for discrimination and outcome prediction of pulmonary embolism: Original Research Article. Emerg Radiol 2024;31:303-11. [Crossref] [PubMed]
17. Hemalakshmi GR, Murugappan M, Sikkandar MY, Santhi D, Prakash NB, Mohanarathinam A. PE-Ynet: a novel attention-based multi-task model for pulmonary embolism detection using CT pulmonary angiography (CTPA) scan images. Phys Eng Sci Med 2024;47:863-80. [Crossref] [PubMed]
18. Lanza E, Ammirabile A, Francone M. nnU-Net-based deep-learning for pulmonary embolism: detection, clot volume quantification, and severity correlation in the RSPECT dataset. Eur J Radiol 2024;177:111592. [Crossref] [PubMed]
19. Xi L, Xu F, Kang H, Deng M, Xu W, Wang D, Zhang Y, Xie W, Zhang R, Liu M, Zhai Z, Wang C. Clot ratio, new clot burden score with deep learning, correlates with the risk stratification of patients with acute pulmonary embolism. Quant Imaging Med Surg 2024;14:86-97. [Crossref] [PubMed]
20. Qiao Y, Gao Y, Chen Y, Ye X, Yan C, Zeng M. Quantitative assessment and risk stratification of random acute pulmonary embolism cases using a deep learning model based on computed tomography pulmonary angiography images. Quant Imaging Med Surg 2025;15:1950-62. [Crossref] [PubMed]
21. Zhang H, Cheng Y, Chen Z, Cong X, Kang H, Zhang R, Guo X, Liu M. Clot burden of acute pulmonary thromboembolism: comparison of two deep learning algorithms, Qanadli score, and Mastora score. Quant Imaging Med Surg 2022;12:66-79. [Crossref] [PubMed]
22. Masutani Y, MacMahon H, Doi K. Computerized detection of pulmonary embolism in spiral CT angiography based on volumetric image analysis. IEEE Trans Med Imaging 2002;21:1517-23. [Crossref] [PubMed]
23. Bouma H, Sonnemans JJ, Vilanova A, Gerritsen FA. Automatic detection of pulmonary embolism in CTA images. IEEE Trans Med Imaging 2009;28:1223-30. [Crossref] [PubMed]
24. Lu NH, Wang CY, Liu KY, Huang YH, Chen TB. AI-Enhanced Deep Learning Framework for Pulmonary Embolism Detection in CT Angiography. Bioengineering (Basel) 2025;12:1055. [Crossref] [PubMed]
25. Abdelhamid A, El-Ghamry A, Abdelhay EH, Abo-Zahhad MM, Moustafa HE. Improved pulmonary embolism detection in CT pulmonary angiogram scans with hybrid vision transformers and deep learning techniques. Sci Rep 2025;15:31443. [Crossref] [PubMed]
26. Yang X, Lin Y, Su J, Wang X, Li X, Lin J, Cheng KT. A two-stage convolutional neural network for pulmonary embolism detection from CTPA images. IEEE Access 2019;7:84849-57.
27. Huang SC, Kothari T, Banerjee I, Chute C, Ball RL, Borus N, Huang A, Patel BN, Rajpurkar P, Irvin J, Dunnmon J, Bledsoe J, Shpanskaya K, Dhaliwal A, Zamanian R, Ng AY, Lungren MP. PENet-a scalable deep-learning model for automated diagnosis of pulmonary embolism using volumetric CT imaging. NPJ Digit Med 2020;3:61. [Crossref] [PubMed]
28. Yuan H, Shao Y, Liu Z, Wang H. An improved faster R-CNN for pulmonary embolism detection from CTPA images. IEEE Access 2021;9:123417-26.
29. Tajbakhsh N, Shin JY, Gotway MB, Liang J. Computer-aided detection and visualization of pulmonary embolism using a novel, compact, and discriminative image representation. Med Image Anal 2019;58:101541. [Crossref] [PubMed]
30. Xu H, Li H, Xu Q, Zhang Z, Wang P, Li D, Guo L. Automatic detection of pulmonary embolism in computed tomography pulmonary angiography using Scaled-YOLOv4. Med Phys 2023;50:4340-50. [Crossref] [PubMed]
31. Batra K, Xi Y, Al-Hreish KM, Kay FU, Browning T, Baker C, Peshock RM. Detection of Incidental Pulmonary Embolism on Conventional Contrast-Enhanced Chest CT: Comparison of an Artificial Intelligence Algorithm and Clinical Reports. AJR Am J Roentgenol 2022;219:895-902. [Crossref] [PubMed]
32. Batra K, Xi Y, Bhagwat S, Espino A, Peshock RM. Radiologist Worklist Reprioritization Using Artificial Intelligence: Impact on Report Turnaround Times for CTPA Examinations Positive for Acute Pulmonary Embolism. AJR Am J Roentgenol 2023;221:324-33. [Crossref] [PubMed]
33. Shapiro J, Reichard A, Muck PE. New Diagnostic Tools for Pulmonary Embolism Detection. Methodist Debakey Cardiovasc J 2024;20:5-12. [Crossref] [PubMed]
34. Amini E, Hille G, Hürtgen J, Surov A, Saalfeld S. Deep learning-based segmentation of acute pulmonary embolism in cardiac CT images. Int J Comput Assist Radiol Surg 2026;21:367-75. [Crossref] [PubMed]
35. Cano-Espinosa C, Cazorla M, Gonzalez G. Computer aided detection of pulmonary embolism using multi-slice multi-axial segmentation. Appl Sci 2020;10:2945.
36. Liu W, Liu M, Guo X, Zhang P, Zhang L, Zhang R, Kang H, Zhai Z, Tao X, Wan J, Xie S. Evaluation of acute pulmonary embolism and clot burden on CTPA with deep learning. Eur Radiol 2020;30:3567-75. [Crossref] [PubMed]
37. Chen Y, Zou B, Guo Z, Huang Y, Huang Y, Qin F, Li Q, Wang C. SCUNet++: Swin-UNet and CNN bottleneck hybrid architecture with multi-fusion dense skip connection for pulmonary embolism CT image segmentation. In: 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). Piscataway (NJ): IEEE; 2024.
38. Guo J, Liu X, Chen Y, Zhang S, Tao G, Yu H, Zhu H, Lei W, Li H, Wang N. AANet: artery-aware network for pulmonary embolism detection in CTPA images. In: Wang L, Dou Q, Fletcher PT, et al., editors. Medical Image Computing and Computer Assisted Intervention - MICCAI 2022. Cham: Springer; 2022:473-83.
39. Huhtanen H, Nyman M, Mohsen T, Virkki A, Karlsson A, Hirvonen J. Automated detection of pulmonary embolism from CT-angiograms using deep learning. BMC Med Imaging 2022;22:43. [Crossref] [PubMed]
40. Weikert T, Winkel DJ, Bremerich J, Stieltjes B, Parmar V, Sauter AW, Sommer G. Automated detection of pulmonary embolism in CT pulmonary angiograms using an AI-powered algorithm. Eur Radiol 2020;30:6545-53. [Crossref] [PubMed]
41. Huang SC, Pareek A, Zamanian R, Banerjee I, Lungren MP. Multimodal fusion with deep neural networks for leveraging CT imaging and electronic health record: a case-study in pulmonary embolism detection. Sci Rep 2020;10:22147. [Crossref] [PubMed]
42. Ma X, Ferguson EC, Jiang X, Savitz SI, Shams S. A multitask deep learning approach for pulmonary embolism detection and identification. Sci Rep 2022;12:13087. [Crossref] [PubMed]
43. Hu Z, Lin HM, Mathur S, Moreland R, Witiw CD, Jimenez-Juan L, Callejas MF, Deva DP, Sejdić E, Colak E. High performance with fewer labels using semi-weakly supervised learning for pulmonary embolism diagnosis. NPJ Digit Med 2025;8:254. [Crossref] [PubMed]
44. Kanne JP, Gotway MB, Thoongsuwan N, Stern EJ. Six cases of acute central pulmonary embolism revealed on unenhanced multidetector CT of the chest. AJR Am J Roentgenol 2003;180:1661-4. [Crossref] [PubMed]
45. Ichinose M, Nomura T, Okada Y, Inagawa H, Sugita M. Quantitative evaluation of the attenuation value of pulmonary thrombus on unenhanced computed tomography. Juntendo Med J 2021;67:338-45.
46. Guo R, Deng M, Xi L, Zhang S, Xu W, Liu M. Chest non‑contrasted computed tomography in detecting acute pulmonary thromboembolism: A single‑center retrospective study. Exp Ther Med 2024;28:304. [Crossref] [PubMed]
47. Chien CH, Shih FC, Chen CY, Chen CH, Wu WL, Mak CW. Unenhanced multidetector computed tomography findings in acute central pulmonary embolism. BMC Med Imaging 2019;19:65. [Crossref] [PubMed]
48. Hagen F, Vorberg L, Thamm F, Ditt H, Maier A, Brendel JM, Ghibes P, Bongers MN, Krumm P, Nikolaou K, Horger M. Improved detection of small pulmonary embolism on unenhanced computed tomography using an artificial intelligence-based algorithm - a single centre retrospective study. Int J Cardiovasc Imaging 2024;40:2293-304. [Crossref] [PubMed]
49. Kim YT, Bak SH, Han SS, Son Y, Park J. Non-contrast CT-based pulmonary embolism detection using GAN-generated synthetic contrast enhancement: Development and validation of an AI framework. Comput Biol Med 2025;198:111109. [Crossref] [PubMed]
50. Pang H, Qi S, Wu Y, Wang M, Li C, Sun Y, Qian W, Tang G, Xu J, Liang Z, Chen R. NCCT-CECT image synthesizers and their application to pulmonary vessel segmentation. Comput Methods Programs Biomed 2023;231:107389. [Crossref] [PubMed]
51. Wu H, Zheng Y, Xia R, Wang L, He Z, Wang X, Guo L. PulmoNet: a transmodal deep learning framework for automated pulmonary embolism diagnosis on non-contrast CT. Biomed Signal Process Control 2026;113:108992.
52. Feng L, Guo L, He W, Ye X, Pan Y, Gao S, Zhang R. DiffCTA: diffusion model-based non-contrast CT angiography with contrast-enhanced mask guidance. Biomed Signal Process Control 2026;112:108645.
53. Kalantar R, Hindocha S, Hunter B, Sharma B, Khan N, Koh DM, Ahmed M, Aboagye EO, Lee RW, Blackledge MD. Non-contrast CT synthesis using patch-based cycle-consistent generative adversarial network (Cycle-GAN) for radiomics and deep learning in the era of COVID-19. Sci Rep 2023;13:10568. [Crossref] [PubMed]
54. Wittram C, Maher MM, Yoo AJ, Kalra MK, Shepard JA, McLoud TC. CT angiography of pulmonary embolism: diagnostic criteria and causes of misdiagnosis. Radiographics 2004;24:1219-38. [Crossref] [PubMed]
55. Schellhaass A, Walther A, Konstantinides S, Böttiger BW. The diagnosis and treatment of acute pulmonary embolism. Dtsch Arztebl Int 2010;107:589-95. [Crossref] [PubMed]
56. Huisman MV, Klok FA. How I diagnose acute pulmonary embolism. Blood 2013;121:4443-8. [Crossref] [PubMed]
57. Qanadli SD, El Hajjam M, Vieillard-Baron A, Joseph T, Mesurolle B, Oliva VL, Barré O, Bruckert F, Dubourg O, Lacombe P. New CT index to quantify arterial obstruction in pulmonary embolism: comparison with angiographic index and echocardiography. AJR Am J Roentgenol 2001;176:1415-20. [Crossref] [PubMed]
58. Braams NJ, Boon GJAM, de Man FS, van Es J, den Exter PL, Kroft LJM, Beenen LFM, Huisman MV, Nossent EJ, Boonstra A, Vonk Noordegraaf A, Ruigrok D, Klok FA, Bogaard HJ, Meijboom LJ. Evolution of CT findings after anticoagulant treatment for acute pulmonary embolism in patients with and without an ultimate diagnosis of chronic thromboembolic pulmonary hypertension. Eur Respir J 2021;58:2100699. [Crossref] [PubMed]
59. Nguyen ET, Hague C, Manos D, Memauri B, Souza C, Taylor J, Dennie C. Canadian Society of Thoracic Radiology/Canadian Association of Radiologists Best Practice Guidance for Investigation of Acute Pulmonary Embolism, Part 1: Acquisition and Safety Considerations. Can Assoc Radiol J 2022;73:203-13. [Crossref] [PubMed]
60. Nguyen ET, Hague C, Manos D, Memauri B, Souza C, Taylor J, Dennie C. Canadian Society of Thoracic Radiology/Canadian Association of Radiologists Best Practice Guidance for Investigation of Acute Pulmonary Embolism, Part 2: Technical Issues and Interpretation Pitfalls. Can Assoc Radiol J 2022;73:214-27. [Crossref] [PubMed]
61. de Jong CMM, Kroft LJM, van Mens TE, Huisman MV, Stöger JL, Klok FA. Modern imaging of acute pulmonary embolism. Thromb Res 2024;238:105-16. [Crossref] [PubMed]
62. Kuzo RS, Levin DL, Bratt AK, Walkoff LA, Suman G, Houghton DE. The use of artificial intelligence to improve detection of acute incidental pulmonary emboli. J Thromb Haemost 2026;24:2473-9. [Crossref] [PubMed]
63. Sorin V, Korfiatis P, Bratt AK, Leiner T, Wald C, Butler C, Cook CJ, Kline TL, Collins JD. Using a Large Language Model for Postdeployment Monitoring of FDA-Approved Artificial Intelligence: Pulmonary Embolism Detection Use Case. J Am Coll Radiol 2025;22:1404-14. [Crossref] [PubMed]
64. Kligerman SJ, Mitchell JW, Sechrist JW, Meeks AK, Galvin JR, White CS. Radiologist Performance in the Detection of Pulmonary Embolism: Features that Favor Correct Interpretation and Risk Factors for Errors. J Thorac Imaging 2018;33:350-7. [Crossref] [PubMed]
65. Abdulaal L, Maiter A, Salehi M, Sharkey M, Alnasser T, Garg P, Rajaram S, Hill C, Johns C, Rothman AMK, Dwivedi K, Kiely DG, Alabed S, Swift AJ. A systematic review of artificial intelligence tools for chronic pulmonary embolism on CT pulmonary angiography. Front Radiol 2024;4:1335349. [Crossref] [PubMed]