Pulmonary embolism (PE) remains a leading cause of cardiovascular mortality worldwide, with an estimated incidence of 60–100 cases per 100,000 persons annually. Prompt diagnosis is critical, as delay in treatment increases mortality risk by as much as 30%. Computed tomography angiography (CTA) of the pulmonary arteries has become the first-line imaging modality for suspected PE, offering high spatial resolution and rapid acquisition. Yet the interpretation of these studies is fraught with challenges: subtle filling defects can be missed, reader variability persists, and the sheer volume of studies in high-throughput settings taxes even the most experienced radiologists.

Into this landscape enters artificial intelligence (AI), not as a replacement for the human reader but as a powerful augmentative tool. Over the past decade, deep-learning algorithms have matured from academic curiosities into clinically deployable systems that can detect pulmonary emboli with sensitivity and specificity rivaling that of subspecialist radiologists. This article examines how AI is enhancing PE detection on CTA, exploring the underlying techniques, clinical benefits, remaining limitations, and future trajectory of this transformative technology.

The Burden of Pulmonary Embolism and the Role of CTA

Pulmonary embolism occurs when a thrombus, typically originating from the deep veins of the lower extremities, travels to the pulmonary arterial circulation and obstructs blood flow. Depending on the clot burden and patient physiology, PE can cause everything from mild dyspnea to sudden cardiac arrest. The clinical presentation is notoriously variable, which is why imaging plays an indispensable role in diagnosis.

CT pulmonary angiography (CTPA) has supplanted ventilation-perfusion scanning as the primary diagnostic test. Modern multi-detector CT scanners acquire sub-millimeter isotropic voxels during a single breath-hold, providing exquisite detail of the pulmonary arteries down to the subsegmental level. However, radiologists must systematically scroll through hundreds of axial images, interpreting axial, coronal, and sagittal reformations while also evaluating the lung parenchyma, mediastinum, and chest wall for alternative or incidental findings. This cognitive load contributes to diagnostic error: systematic reviews report that up to 30% of acute PEs are initially missed on CTA, with false-negative rates highest for subsegmental and central emboli that are subtle or obscured by respiratory motion or streak artifact from contrast material.

The AI Advantage in Medical Imaging

Artificial intelligence applied to radiology has evolved rapidly, driven by advances in deep learning, increases in computational power, and the availability of large, annotated datasets. Convolutional neural networks (CNNs), in particular, excel at extracting hierarchical features from images without the need for handcrafted rules. In the context of PE detection, AI models are trained to identify the visual signature of a filling defect: a low-attenuation region surrounded by contrast-opacified blood, often with a characteristic vessel cutoff or central filling defect morphology.

Unlike traditional computer-aided detection (CAD) systems that rely on manually defined thresholds and heuristics, modern deep-learning approaches learn the statistical patterns of emboli directly from thousands of labeled CTPA studies. These models can generalize across variations in patient anatomy, scanner manufacturer, contrast timing, and image noise, offering robustness that surprises many experienced practitioners. Moreover, AI does not tire, does not skip images, and can work 24/7, making it an ideal safety net for high-volume centers.

Key AI Techniques for Pulmonary Embolism Detection

Convolutional Neural Networks

The backbone of most PE detection algorithms is a convolutional neural network trained on a large repository of CTPA exams. Early models operated on axial slices independently, but modern architectures incorporate three-dimensional context by analyzing volumetric patches or using recurrent neural networks to capture inter-slice dependencies. U-Net and its variants are particularly popular for segmentation tasks, allowing the model to outline the clot precisely and provide a confidence score. For example, the model described in a 2020 Radiology study achieved a sensitivity of over 90% at a low false-positive rate, demonstrating the feasibility of deep learning for PE detection.

Attention Mechanisms and Transformers

More recent innovations include attention-based models that allow the network to focus on the most informative regions of the image, analogous to a radiologist zooming in on a suspicious vessel. Vision transformers, derived from natural language processing architectures, have shown promise in capturing long-range spatial relationships, which may help differentiate true emboli from mimics such as flow artifacts, beam hardening, or lymph nodes. These sophisticated models offer incremental gains in specificity, reducing the number of false-positive alerts and improving radiologist acceptance.

Multi-Task Learning

Many commercial AI systems now incorporate multi-task learning, simultaneously performing clot segmentation, laterality assignment, and even quantification of clot burden (e.g., right ventricular dysfunction assessment via the right ventricle-to-left ventricle ratio). This multi-output approach provides a comprehensive interpretive aid that goes beyond simple detection, helping clinicians triage patients and guide treatment decisions such as thrombolysis versus anticoagulation alone.

Clinical Impact of AI-Enhanced PE Detection

The integration of AI into the CTPA workflow has yielded measurable benefits across multiple dimensions:

Improved Sensitivity and Reduced False Negatives

Several multi-center studies have demonstrated that AI-based CAD systems detect a higher proportion of PEs than unaided radiologists, particularly for subsegmental emboli and subtle central emboli. A 2022 systematic review and meta-analysis in JAMA Network Open found that deep-learning models achieved a pooled sensitivity of 92% (95% CI, 87–95%), compared with 80% for radiologists reading without AI assistance. The reduction in false negatives means fewer patients are sent home with untreated PE, potentially preventing morbidity and mortality.

Accelerated Read Times and Workflow Efficiency

AI algorithms can process a CTPA study in seconds, flagging candidate emboli before the radiologist even opens the case. When integrated into the picture archiving and communication system (PACS), this results in a dramatic reduction in interpretation time. One study reported a 40% decrease in average reading time for PE studies when using an AI overlay. This time savings can be redeployed to more complex cases, to reduce burnout, or to shorten report turnaround times, which is especially valuable in emergency departments and on-call settings.

Standardization Across Readers and Institutions

Practitioner variability is a well-documented problem in radiology. A senior thoracic radiologist with years of experience may detect PE with high specificity, while a junior resident or a general radiologist covering nights may have a lower threshold for calling a filling defect. AI provides a consistent second reader that applies the same detection criteria to every case, regardless of time of day or user expertise. This standardization helps maintain diagnostic quality across shifts and between different hospitals within a health system.

AI Integration: Challenges and Considerations

Despite its promise, the deployment of AI for PE detection is not without obstacles. Understanding these limitations is essential for safe and effective adoption.

Data Dependency and Generalizability

Deep-learning models are only as good as the data on which they are trained. Many existing PE detection algorithms have been trained on datasets from large academic medical centers using specific scanner models and contrast protocols. When deployed in community hospitals with different equipment or injection protocols, performance can degrade. The phenomenon of domain shift means that a model that achieves 95% sensitivity in one institution may drop to 80% in another. Rigorous multi-institutional validation and continuous retraining with local data are necessary to ensure robust performance.

False Positives and Operator Trust

AI models can generate false-positive alerts—areas that appear suggestive of PE but are actually caused by flow artifacts, partial volume averaging, or respiratory motion. If the false-positive rate is too high, radiologists may develop “alert fatigue,” dismissing the AI’s findings and missing true positives. Manufacturers are working to reduce false positives through improved training with adversarial examples and by incorporating context-aware post-processing (e.g., rejecting findings in vessels smaller than a certain caliber). Nonetheless, the final interpretation must remain the radiologist’s judgment.

Regulatory and Ethical Hurdles

Commercial AI products for PE detection must receive clearance from bodies such as the U.S. Food and Drug Administration (FDA). As of early 2025, several algorithms have obtained 510(k) clearance, including products from vendors like Siemens Healthineers, GE Healthcare, and Imbio. However, the regulatory landscape is evolving. Questions about liability when an algorithm misses a finding, data privacy for AI training sets, and the need for post-market surveillance are all active topics of debate. Institutions must establish governance frameworks that define how AI outputs are used, documented, and audited.

The Future of AI in PE Detection

Looking ahead, the role of AI in PE detection is likely to expand in several dimensions:

Beyond Detection to Prediction and Prognostication

Future AI systems may not only detect emboli but also predict clinical outcomes. By combining imaging features with electronic health record data (e.g., vital signs, lab values, comorbidities), deep-learning models could estimate the risk of death, bleeding, or recurrence, enabling personalized treatment decisions. Early work in this area, such as the multimodal model from Stanford, shows that integrating clinical variables with imaging features improves prognostication over imaging alone.

Real-Time AI at the Scanner

Another promising frontier is the deployment of AI directly on the CT scanner console. This would allow the algorithm to analyze images as they are reconstructed, immediately notifying the technologist if the study is inadequate (e.g., poor contrast timing) or if a high-probability PE is identified. Such real-time feedback could trigger appropriate protocol modifications or even prioritize the study in the reading queue, shaving critical minutes off the diagnosis-to-treatment interval.

Integration with Natural Language Processing

AI systems that combine vision and language are emerging. These could automatically generate structured reports containing the location, extent, and morphology of detected PEs, as well as associated findings such as right heart strain. The radiologist would then only need to verify the machine-generated content, further reducing cognitive burden and improving report consistency.

Conclusion

Artificial intelligence is no longer a futuristic concept in radiology; it is a practical tool that is already enhancing the detection of pulmonary emboli on CT angiography. By providing a consistent, tireless, and rapid second read, AI helps radiologists catch subtle clots, reduces interpretation time, and standardizes care across institutions. However, successful implementation requires careful attention to model generalizability, false-positive management, and regulatory compliance. The optimal outcome is a collaborative human–AI partnership where the radiologist remains the final arbiter, empowered rather than replaced by the technology. As algorithms continue to improve and integrate more deeply into the clinical workflow, the promise of earlier, more accurate diagnosis of pulmonary embolism will translate into lives saved and better outcomes for patients worldwide.