Automated detection of pulmonary embolism from CT-angiograms using deep learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Pulmonary embolism (PE) diagnosis relies on computed tomography pulmonary angiograms (CTPAs).
• Automated detection of PE can improve diagnostic efficiency and accuracy.
• Developing AI models for medical image analysis often requires large, meticulously annotated datasets.

Purpose of the Study:

• To develop and evaluate a deep neural network for automated PE detection from CTPAs.
• To assess the model's performance using only weakly labeled training data.
• To compare two model versions pre-trained on different image datasets.

Main Methods:

• A deep neural network combining InceptionResNet V2 and LSTM was developed.
• Two model versions were trained using weakly labeled CTPAs, pre-trained on chest X-rays (Model A) or natural images (Model B).
• Model performance was evaluated using ROC curves, precision-recall curves, sensitivity, specificity, and predictive values.

Main Results:

• Both models demonstrated strong performance on stack- and slice-based PE detection.
• Model A (pre-trained on X-rays) slightly outperformed Model B, achieving 93.5% specificity and 86.6% sensitivity.
• ROC AUC scores were high for both models (0.94 vs 0.91), with no statistically significant difference.

Conclusions:

• Deep learning models can achieve excellent performance in PE detection from CTPAs.
• Effective PE detection is possible even with relatively small, weakly annotated datasets.
• This approach shows promise for improving automated diagnostic tools in radiology.