A Deep Learning Model for Diagnosis of Pulmonary Embolism from CT Pulmonary Angiography

Updated: Sep 21

Yash Pargaonkar Background:

Clotting of blood is a normal physiological phenomenon to prevent excessive bleeding from an injured blood vessel. The human body is usually able to form and break down clots inside blood vessels. However, when the body is unable to dissolve a clot inside the blood vessel, the clot expands to form a ‘thrombus’. Although legs are the most common site for the occurrence of thrombosis, they can form anywhere in the body1. When a thrombus gets dislodged and transported to another region in the body through blood, it is called an embolus. Pulmonary embolism is a disorder affecting the pulmonary vasculature. Pulmonary embolism occurs due to obstruction of blood vessels supplying blood to the lungs. The obstruction prevents the transport of blood to the lungs for effective oxygenation. As a consequence, the body cannot receive oxygen. Moreover, increased obstruction of pulmonary blood vessels leads to increased strain on the heart. It may result in a spectrum of hemodynamic consequences from asymptomatic to life-threatening medical emergencies. Pulmonary embolism is the third most common vascular cause of death after myocardial infarction (heart attack) and stroke2. The sudden onset of symptoms and the high mortality rates make this disorder difficult to manage.

Key Statistics:

The annual incidence of pulmonary embolism is approximately 60 to 70 cases per 1,00,000 people, resulting in approximately 1,00,000 to 3,00,000 deaths annually2.

The risk factors for pulmonary embolism include disorders of blood clotting, surgery, prolonged immobility, old age, cancer, fracture of long bones, and many more2. The clinical features of massive pulmonary embolism include rapidly progressive difficulty in breathing, increased heart rate, palpitations, chest pain, anxiety, falling blood pressure, and sweating3.

Role of Medical Imaging:

Diagnosing pulmonary embolism requires a high clinical suspicion as the diagnostic tests are specialized for the same. D-dimer assay and CT Pulmonary Angiography (CTPA) are the two tests used to diagnose pulmonary embolism. CTPA is a radiological procedure used to obtain images of pulmonary vessels that originate from the heart and go to the lungs4. If present, the embolus is visible on CTPA. Other classical tests like ventilation-perfusion scan and invasive pulmonary angiography are used less commonly as CTPA has now become the investigation of choice for diagnosis of pulmonary embolism5. The treatment for pulmonary embolism involves cardiopulmonary support, anticoagulation therapy, and reperfusion of the pulmonary artery6. However, the duration of anticoagulation after the onset of symptoms plays an important part in determining the mortality of the disease. Hence, there is a need for rapid identification of PE to enable optimal medical care. Therefore, we developed a Deep Learning (DL) model called “DxPE AI Screen” to aid radiologists/physicians in the faster diagnosis of pulmonary embolism.

The Validation Study:

The objective of our study was to externally validate the performance of “DxPE AI Screen” to detect pulmonary embolism on CTPA using relevant data from the CTPAs performed at a tertiary care hospital. The diagnosis of PE included both clinical information (D-dimer test, echocardiographic findings) and CTPA findings. The external test dataset consisted of 251 CTPAs (55 positive and 196 negative for PE) performed between January 2018 and January 2021. The expertise of three radiologists with 23, 15, and 9 years of experience, respectively, was used to classify scans and highlight emboli, if the scans were found to be positive for PE. The radiologists used several indications to identify pulmonary embolism, which included complete filling defects, ‘polo mint sign’, ‘railway track sign’, and peripherally located defects7,8,9.

Figure 1: Example of model prediction on a CT slice with (a) embolism present and (b) no embolism.

AI Model:

Because the model predicted PE only at the slice level, we developed an approach to predict PE at the scan level. Model training was monitored by its performance on separate validation sets and the model with the lowest validation loss was selected as the final model. The test set was evaluated for sensitivity, specificity, positive predictive value, negative predictive value, accuracy, F1 score, and Area under the Receiver Operating Characteristic curve. The model correctly identified 44 out of 55 scans with pulmonary embolism.

Amongst 251 patients, 145 were males and 106 were females with a mean age of 49.7 ± 17.4 years. The model was able to mark the presence of pulmonary embolism in 44 of the 55 positive cases. The similarity between the reference standard and AI-predicted embolus mask was high. The model required a mean time of 30.15 ± 11.03 seconds to scan the CTPA and categorize it. Table 1 shows the performance of U-Net segmentation model. The model obtained high sensitivity, ROC, and NPV in identifying PE. The AUC obtained by the model at the scan level was comparable to AUCs reported in previous studies.


Table 1: Performance of U-Net segmentation model

Fig 2: The Area Under the ROC Curve obtained by the DL model.

“DxPE AI Screen” was trained and tested on a larger sample than most of the previous studies. Since the prevalence of scans that were positive for PE was low, the model obtained a low positive predictive value on the external test dataset. Additionally, the choice of selecting a particularly high sensitivity operating point also contributed to low PPV and high NPV. While this approach could falsely label a negative scan as a positive one, it is important to understand that this model is built for suspected PE scans to be further interpreted by a physician/radiologist. Due to the high mortality rate of PE, it is more important for all the suspected cases to be correctly identified even at the expense of a few false positive outcomes.

Future Implications:

With the increasing use of CTPAs in hospitals, a timely and accurate diagnosis is a key step to supporting radiologists and physicians in triage and a quick review of such suspect cases. A well-trained AI model can help reduce this burden by identifying the positive CTPA scans for the radiologist to assess.

The full article is available at: https://assets.researchsquare.com/files/rs-1909034/v1/8558955e-436a-4969-bc8a-20ae7a187e03.pdf?c=1659743562


  1. Yoshimura, N., Hori, Y., Horii, Y. et al. Where is the most common site of DVT? Evaluation by CT venography. Jpn J Radiol 30, 393–397 (2012). https://doi.org/10.1007/s11604-012-0059-6

  2. Bĕlohlávek J, Dytrych V, Linhart A. Pulmonary embolism, part I: Epidemiology, risk factors and risk stratification, pathophysiology, clinical presentation, diagnosis, and non-thrombotic pulmonary embolism. Exp Clin Cardiol. 2013;18(2):129-138.

  3. Morrone D, Morrone V. Acute Pulmonary Embolism: Focus on the Clinical Picture [published correction appears in Korean Circ J. 2018 Jul;48(7):661-663]. Korean Circ J. 2018;48(5):365-381. doi:10.4070/kcj.2017.0314

  4. Marshall PS, Mathews KS, Siegel MD. Diagnosis and management of life-threatening pulmonary embolism. J Intensive Care Med. 2011;26(5):275-294. doi:10.1177/0885066610392658

  5. Anderson DR, Kahn SR, Rodger MA, et al. Computed Tomographic Pulmonary Angiography vs Ventilation-Perfusion Lung Scanning in Patients With Suspected Pulmonary Embolism: A Randomized Controlled Trial. JAMA. 2007;298(23):2743–2753. doi:10.1001/jama.298.23.2743

  6. Martinez Licha CR, McCurdy CM, Maldonado SM, Lee LS. Current Management of Acute Pulmonary Embolism. Ann Thorac Cardiovasc Surg. 2020;26(2):65-71. doi:10.5761/atcs.ra.19-00158

  7. Coche EE, Muller NL, Kim K, et al. Acute pulmonary embolism: ancillary findings at spiral CT. Radiology 1998; 207:753-758.

  8. Gaye B, Remy J, Remy-Jardin M. Non-traumatic thoracic emergencies: CT diagnosis of acute pulmonary embolism—the first 10 years. Eur Radiol 2002; 12:1886-1905.

  9. https://pubs.rsna.org/doi/10.1148/rg.245045008

25 views0 comments