Deep Learning-Based Algorithm for Automatic Detection of Pulmonary Embolism in Chest CT Angiograms.

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Prompt recognition and treatment of pulmonary embolism (PE) are critical for reducing mortality.
• Chest computed tomography angiography (CTA) is a primary imaging modality for PE diagnosis.
• Deep learning (DL) offers potential for automating image analysis in radiology.

Purpose of the Study:

• To validate a DL-based algorithm for automated detection of PE on chest CTAs.
• To assess the diagnostic performance of the DL algorithm in a real-world setting.
• To evaluate the algorithm's ability to alert radiologists for urgent interpretation of potential PE cases.

Main Methods:

• A hybrid 3D/2D UNet convolutional neural network (CNN) topology was employed.
• The algorithm was trained on diverse datasets considering scanner vendors, patient age, and image acquisition parameters.
• Validation was performed on 387 anonymized chest CTAs from multiple clinical sites, with ground truth established by three radiologists.

Main Results:

• The DL algorithm demonstrated a sensitivity of 91.4% and specificity of 91.5% for PE detection.
• Overall accuracy was 91.5%, with a balanced performance in identifying positive and negative PE cases.
• False negatives were associated with chronic or subsegmental PEs, while false positives stemmed from artifacts and anatomical variations.

Conclusions:

• The developed DL algorithm exhibits high diagnostic accuracy for PE detection on CTAs.
• The algorithm shows balanced sensitivity and specificity, making it a promising tool for clinical use.
• This technology can support radiologists in prioritizing urgent interpretations, potentially enhancing patient care for PE.