The Role of AI in Pulmonary Embolism Detection

Pulmonary Embolism (PE) is one of the leading causes of morbidity and mortality worldwide. Factors contributing to the development of PE can be summarized by Virchow’s triad, which includes hypercoagulability, endothelial injury, and stasis of blood flow. Risk factors for PE include malignancy, surgery, prolonged immobilization, obesity, estrogen replacement therapy and pregnancy. Diagnosis of PE by CT Pulmonary Angiogram (CTPA) is the current gold standard. Patients with PE require rapid diagnosis, and management to improve clinical outcomes. In recent years, AI algorithms have been introduced to detect pulmonary emboli, in an aim to improve diagnostic efficiency, patient outcomes, and workflow. In this blog post, we will discuss two studies that assess the accuracy of AI algorithms in detecting PE.

1- Accuracy of a deep neural network for automated pulmonary embolism detection on dedicated CT pulmonary angiograms

This study published in The European Journal of Radiology aimed to assess the performance of a Deep Neural Network (DNN)-based algorithm for automated PE detection on CTPA scans. This was a retrospective study that included 903 patients from a single center who had CTPA scans due to suspicion of PE between September 2022 to January 2023.

The AI evaluation of CTPA studies was performed. Subgroup analyses were conducted to assess the algorithm’s performance considering body mass index (BMI) and PE location. BMI-based analysis was performed in obese and non-obese patients by dividing patients into two subgroups, BMI < 30 kg/m² and BMI ≥ 30 kg/m². PE location-based analysis was done by dividing the pulmonary artery system into three levels:

1- Central arteries (main pulmonary artery, right and left pulmonary arteries).

2- Lobar branches (right upper lobe, right middle lobe, right lower lobe, left upper lobe, lingula, and left lower lobe).

3- Peripheral regions (segmental and subsegmental branches).

In cases with multiple emboli, the location of the most proximal PE was considered.

Results

The prevalence of PE was 12.2 % (n = 110). The model achieved a sensitivity of 84.6%, specificity of 95.1%, PPV of 70.5%, and NPV 97.8% and had an overall accuracy of 93.8%. Among the false positive cases (n = 39), common sources of error included lung masses, pneumonia, and contrast flow artifacts. Among the false negative cases (n = 17) causes included chronic and subsegmental PEs.

The prevalence of PE was similar in obese (12.2%) and non-obese patients (11.9%) in this cohort. In the obese and non-obese groups, the algorithm’s accuracy for detecting PE was 92.5% and 94.6% respectively.

The PEs in this cohort included 21 central, 30 lobar, and 59 peripheral cases of PE. In the central arteries, all emboli were detected. In the lobar branches, 29 of 30 emboli were detected, while in the peripheral regions, 43 of 59 emboli were correctly detected. Missed PEs were identified as chronic subsegmental PEs and acute/chronic PEs in segmental or subsegmental arteries by further expert reading.

Key Findings

• This AI algorithm could detect all central PEs, most lobar PEs and fewer peripheral PEs. This points to excellent diagnostic accuracy of central PEs, but further improvement is still required for lobar and peripheral PE detection.
• No significant difference between BMI subgroups.
• Algorithms such as this one could potentially aid radiologists in exam prioritization and rapid diagnosis of PE.

2. Implementation of an AI Algorithm in Clinical Practice to Reduce Missed Incidental Pulmonary Embolisms on Chest CT and Its Impact on Short-Term Survival

This study, published in Investigative Radiology, was an observational single-center study that included 5,298 chest CT scans performed for indications other than suspected pulmonary embolism (PE). There were 2 cohorts in this study. Cohort 1 consisted of 1,964 patients whose radiology reports were produced before the application of the FDA approved AI algorithm for incidental PE detection (Aidoc Medical), and cohort 2 consisted of 3,334 patients whose CT scans were assessed after the implementation of the AI algorithm (Aidoc Medical). For comparison, cohort 1 was reviewed retrospectively by the AI algorithm.

In both cohorts, any differences between the original radiology reports and the AI output were reviewed by 2 cardiothoracic radiologists.

Results:

In cohort 1 the prevalence of confirmed incidental PE was 2.2% (n = 42), and the AI detected 61 suspicious incidental PEs, resulting in a sensitivity of 95%, a specificity of 99%, a PPV of 69%, and an NPV of 99%. In cohort 1, radiologists overlooked 50% of incidental PE cases.

In cohort 2, the prevalence of confirmed incidental PEs was 1.7% (56/3334), and the AI detected 59 suspicious cases with a sensitivity of 90%, a specificity of 99%, a PPV of 95%, and a NPV of 99%. The rate of incidental PEs missed by radiologists dropped to 7.1% after AI implementation, showing substantial improvement.

Key Findings

• The implementation of an AI algorithm significantly reduced the rate of missed incidental PEs from 50% to 7.1%, thereby enhancing diagnostic accuracy. Despite this improvement, the 90-day mortality rate remained unchanged possibly due to the fact that most patients in this cohort that had incidental PEs were on anticoagulant therapy.
• These findings highlight the AI tool’s potential to assist radiologists in accurately identifying incidental PEs.

In conclusion, AI implementation has the potential to improve diagnostic accuracy and speed in PE detection, which can help in prioritizing critically ill patients. However, further improvement is still required and larger multi-center studies need to assess the generalizability of algorithm performance.