Enhanced classification performance using deep learning based segmentation for pulmonary embolism detection in CT...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Pulmonary embolism (PE) is a critical condition requiring timely diagnosis.
• CT pulmonary angiography (CTPA) is a primary imaging modality for PE detection.
• Accurate and automated diagnostic tools are needed to improve efficiency and patient outcomes.

Purpose of the Study:

• To develop and evaluate a deep learning-based algorithm for automated PE classification in CTPA examinations.
• To achieve high accuracy in distinguishing between patients with and without PE.

Main Methods:

• Utilized the nnU-Net deep learning framework for segmentation.
• Trained the model on 700 CTPA examinations from a single institution.
• Applied logical rules based on PE volume and probability thresholds for classification.
• Validated the model on two independent external datasets.

Main Results:

• Achieved high sensitivity (96.1%) and specificity (94.6%) in internal testing.
• Demonstrated strong performance on external datasets with AUROCs of 98.6% and 98.5%.
• The algorithm correctly classified a large majority of PE and non-PE cases across all evaluated datasets.

Conclusions:

• The developed automatic pipeline using nnU-Net achieves state-of-the-art diagnostic performance for PE detection in CTPA.
• The deep learning approach offers a promising tool for accurate and efficient PE classification.
• This AI-driven method has the potential to enhance clinical decision-making in diagnosing pulmonary embolism.