Pulmonary embolism (PE) remains a leading cause of cardiovascular mortality worldwide, with an estimated incidence of 60–70 cases per 100,000 population annually. This life-threatening condition occurs when a blood clot—typically originating from deep veins in the legs—migrates to the lungs, obstructing pulmonary arteries and impairing gas exchange. Rapid and accurate diagnosis is essential to initiate anticoagulation or thrombolytic therapy, reduce morbidity, and prevent fatal outcomes. Computed tomography angiography (CTA) has become the standard of care for PE diagnosis, offering high spatial resolution and rapid acquisition. However, interpreting CTA studies for subtle or peripheral emboli can be challenging, even for experienced radiologists. Image processing techniques have emerged as a critical tool to augment the diagnostic capability of CTA, enabling earlier detection, reducing interpretive variability, and ultimately improving patient outcomes.

Role of Image Processing in CTA for PE

Image processing encompasses a broad range of computational techniques applied to raw image data to enhance features, suppress artifacts, and extract clinically relevant information. In the context of CTA for PE detection, these methods work synergistically to improve visualization of pulmonary vasculature, illuminate occlusive and non-occlusive clots, and automate the detection of suspicious regions. The integration of image processing into clinical workflows has shifted the paradigm from purely manual interpretation to a hybrid human–computer diagnostic process, where algorithms serve as a second reader or triage tool.

Enhancement Techniques

Preprocessing steps are foundational to effective image analysis. Contrast enhancement methods, such as adaptive histogram equalization or CLAHE (contrast limited adaptive histogram equalization), adjust local intensity distributions to improve the conspicuity of the pulmonary arteries relative to surrounding lung parenchyma. Edge detection algorithms, including Sobel, Canny, or Laplacian of Gaussian filters, delineate vessel boundaries and can highlight abrupt changes in attenuation that may represent intravascular filling defects. Denoising filters (e.g., bilateral filters, non-local means, or deep learning–based denoisers) reduce quantum noise inherent in low-dose CTA protocols, which are increasingly used to limit radiation exposure. These enhancement techniques not only aid human readers but also prepare images for subsequent automated analysis by ensuring consistent image quality across diverse scanner platforms and acquisition parameters.

Segmentation and Vessel Extraction

Accurate delineation of the pulmonary arterial tree is a prerequisite for reliable PE detection. Segmentation algorithms—ranging from region growing and level sets to deep convolutional neural networks (CNNs)—extract the pulmonary vasculature from surrounding lung, mediastinal structures, and contrast-enhanced thoracic vessels. Once segmented, the vessel lumen can be analyzed for intraluminal filling defects. Hierarchical segmentation approaches, such as vesselness filters (e.g., Frangi filter), exploit the tubular geometry of arteries to identify candidate regions and separate arteries from veins. Advanced U-Net architectures and their variants (e.g., V-Net, Attention U-Net) have demonstrated excellent performance in segmenting subsegmental pulmonary arteries, which are notoriously difficult to assess manually. Image processing enables reproducible quantification of vascular morphology, including branching patterns and cross-sectional area, which can be correlated with clot burden.

Automated Detection Algorithms

Machine learning and deep learning have revolutionized automated PE detection. Traditional machine learning approaches relied on handcrafted features—intensity, shape, texture—extracted from candidate regions and classified using random forests or support vector machines. These systems achieved reasonable sensitivity but suffered from high false-positive rates due to complex anatomical variants and imaging artifacts. Deep learning models, particularly CNNs, now dominate the field. A typical pipeline involves: (1) segmentation of pulmonary arteries, (2) sliding-window or region proposal generation, (3) classification of each candidate as embolus or non-embolus using a pretrained CNN (e.g., ResNet, DenseNet), and (4) postprocessing to aggregate detections and reduce false positives. Recent publications have reported per‑embolus sensitivity exceeding 90% with a false‑positive rate below 1 per patient, rivaling expert radiologist performance. Some systems incorporate temporal information from contrast enhancement dynamics or leverage 3D convolutional networks (C3D, I3D) to capture volumetric context.

Deep Learning Models in Clinical Validation

Several commercial AI solutions for PE detection have received FDA clearance, including Aidoc’s PE algorithm and Viz.ai’s PE triage tool. These systems operate in real time on CTA source images, flagging suspicious studies for immediate review. Large‑scale retrospective and prospective studies have validated their impact: reduced turnaround times from minutes to seconds, increased detection of incidental PEs in emergency department populations, and improved inter‑reader agreement. For instance, a 2022 multicenter trial published in Radiology demonstrated that an AI‑assisted triage protocol decreased median time to anticoagulation by 42% compared to standard workflow. Such evidence underscores the translational value of image processing in PE care.

Quantification and Severity Assessment

Beyond binary detection, image processing enables objective quantification of clot burden—a key predictor of right ventricular (RV) dysfunction and early mortality. Parameters such as the Qanadli index, Mastora score, or volumetric embolic load can be automatically computed from segmented thrombus regions. These metrics correlate with RV/left ventricular (LV) diameter ratios on axial or four‑chamber views, which are themselves measurable via image processing algorithms. Automated calculation of RV/LV ratio, pulmonary artery diameter, and interventricular septal bowing provides a composite risk stratification that can be integrated into clinical decision support systems. By standardizing severity assessment, image processing reduces subjectivity and facilitates longitudinal monitoring of patients undergoing thrombolytic or surgical intervention.

Benefits of Image Processing in PE Detection

The adoption of image processing techniques into clinical practice yields tangible benefits across multiple dimensions of patient care and radiology workflow. The following list summarizes the most significant advantages:

  • Increased Diagnostic Accuracy: Enhancement and automated detection methods improve identification of small, subsegmental emboli that are frequently overlooked in routine interpretation. A meta‑analysis of deep learning–based systems reported pooled sensitivity of 92% and specificity of 91% for detecting acute PE on CTA.
  • Faster Diagnosis and Triage: AI‑based triage tools prioritize studies with suspected PE, notifying radiologists immediately. In busy emergency departments, this reduces mean time to report from several hours to under 15 minutes, expediting life‑saving interventions.
  • Reduced Inter‑observer Variability: Image analysis algorithms apply consistent criteria across all cases, minimizing the impact of reader fatigue, experience level, or interpretive bias. Several studies have shown that AI assistance improves agreement between junior and senior radiologists from fair (κ=0.45) to good (κ=0.78).
  • Improved Workflow Efficiency: Automated segmentation of the pulmonary arteries reduces time spent on manual windowing and scroll‑through examination. Radiologists can allocate cognitive resources to complex cases or incidental findings.
  • Enhanced Severity Stratification: Quantitative metrics derived from image processing—embolus volume, RV/LV ratio, pulmonary artery diameter—provide objective data for risk‑adjusted management, such as selecting patients for catheter‑directed thrombolysis or surgical embolectomy.
  • Better Patient Outcomes: Early and accurate detection, coupled with timely treatment, lowers the risk of hemodynamic collapse, chronic thromboembolic pulmonary hypertension, and death. Population‑level studies suggest that widespread adoption of AI‑assisted PE detection could prevent thousands of misdiagnoses annually.

Challenges and Current Limitations

Despite the remarkable progress, several barriers impede the full integration of image processing into routine PE practice. Addressing these challenges is essential for safe, equitable, and effective deployment.

Data Heterogeneity

CTA images come from different CT manufacturers (Siemens, GE, Philips, Canon), each with distinct reconstruction kernels, dose levels, and contrast injection protocols. Algorithms trained on a single institution’s data often degrade in performance when applied to external datasets. Domain adaptation and data augmentation techniques—such as simulating different scanner noise profiles—mitigate this issue only partially. Large‑scale, multicenter training sets with diverse demographics and acquisition parameters are needed to ensure robustness.

Annotation and Ground Truth

Supervised deep learning relies on large volumes of accurately labeled training data. Creating voxel‑wise annotations for PE—a task that requires expert radiologists and can take 30–60 minutes per CTA volume—is resource‑intensive. The resulting labels may still contain inter‑observer discrepancies, especially for small or chronic emboli. Weakly supervised approaches and semi‑supervised learning are active research areas, but they have not yet matched the performance of fully supervised methods.

Interpretability and Trust

Many radiologists remain skeptical of “black‑box” AI systems. For image processing to be accepted, physicians must understand why an algorithm flagged a particular region. Explainable AI techniques—such as saliency maps, class activation maps (Grad‑CAM), or feature visualization—are increasingly incorporated into commercial products. However, these explanations can be misleading or incomplete. Building clinician trust requires transparent, validated models and continuous education about the strengths and limitations of the technology.

Regulatory and Reimbursement Hurdles

Obtaining FDA clearance or CE marking for AI‑based PE detection tools is a rigorous process requiring evidence of safety, efficacy, and clinical benefit. Even with clearance, widespread adoption depends on reimbursement models. In the United States, Current Procedural Terminology (CPT) codes for computer‑aided detection exist but may not adequately cover the integration, maintenance, and supervision costs of AI systems. Radiologists worry about liability when relying on algorithm outputs, and clear medicolegal frameworks are still evolving.

Incidental Findings and Overdependence

AI systems optimized for PE detection may overlook other critical findings, such as aortic dissection, pulmonary nodules, or coronary artery disease. Radiologists must remain vigilant to avoid “automation bias,” where they over‑rely on the algorithm and miss diagnoses. Workflow integration that presents AI results as a second opinion rather than a definitive reading is recommended. Additionally, false‑positive alerts can contribute to alarm fatigue, especially in high‑volume emergency departments.

Future Directions

Ongoing research promises to overcome current limitations and extend the capabilities of image processing in PE diagnosis. The following directions are particularly promising:

Multimodal AI Integration

Combining CTA image data with clinical information (D‑dimer, Wells score, vital signs) and other imaging modalities (ventilation/perfusion scans, echocardiography) through multimodal deep learning frameworks can improve diagnostic accuracy and risk stratification. Early fusion, late fusion, or attention‑based models that learn cross‑modal correlations are under investigation. Such systems could provide a probabilistic diagnosis that incorporates pretest probability, imaging findings, and hemodynamic context.

Real‑Time Processing and Point‑of‑Care

Advances in GPU computation and model compression enable real‑time inference on CT consoles or cloud‑based platforms. Future systems may perform “live” PE detection during the scan acquisition, allowing technologists to immediately notify the radiologist of a positive study. Portable CT scanners equipped with on‑device AI could extend the benefits to resource‑limited or rural settings where subspecialty radiology access is scarce.

Federated Learning and Privacy Preservation

Training robust models requires data from many institutions, yet sharing patient data raises privacy and regulatory concerns. Federated learning trains a global model across decentralized data without transferring raw images. Several consortia (e.g., the Medical Imaging and Data Resource Center, MIDRC) are exploring federated frameworks for PE detection. Early results indicate that federated models can approach centralized performance while preserving data sovereignty.

Explainable and Causal AI

Next‑generation image processing algorithms will incorporate causal reasoning—for instance, modeling how a filling defect arises from a clot rather than a breathing artifact—and provide counterfactual explanations (e.g., “if this region were removed, the PE likelihood would drop from 0.8 to 0.2”). Such approaches could boost clinician trust and facilitate regulatory approval by clarifying decision rationale.

Expansion to Chronic and Subclinical PE

Image processing techniques are being refined to detect chronic thromboembolic disease, which presents with webs, stenoses, and mosaic perfusion patterns on CTA. Differentiating acute from chronic PE guides management, since chronic cases may require pulmonary endarterectomy rather than anticoagulation. Moreover, as CTA becomes more sensitive, subclinical PE in conditions like COVID‑19 or malignancy could be systematically evaluated, potentially altering surveillance protocols.

Conclusion

Image processing has transformed the detection of pulmonary embolism in CT angiography, evolving from basic enhancement to sophisticated, AI‑driven diagnostic systems. These techniques enhance visualization, automate detection, quantify clot burden, and improve risk stratification, leading to faster and more accurate diagnosis. Although challenges related to data heterogeneity, annotation, interpretability, and regulatory approval remain, ongoing advances in multimodal AI, real‑time processing, federated learning, and explainable algorithms promise to further elevate the standard of care. As image processing continues to mature, its integration into clinical workflows will not only improve outcomes for individual patients but also streamline radiology operations and reduce the global burden of missed PE. The synergy between human expertise and computational power represents the future of precision medicine in thoracic imaging.