Automated AI detection of thoracic aortic dissection on CT imaging.

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Aortic dissection (AD) is a critical medical emergency requiring prompt diagnosis.
• Current diagnostic methods can be time-consuming and may not always be definitive.
• Automated detection of AD can significantly improve patient outcomes.

Purpose of the Study:

• To develop and validate an artificial intelligence (AI) algorithm for automated detection and sub-classification of aortic dissection (AD).
• To assess the performance of the AI algorithm on both internal and external, heterogeneous datasets.
• To provide a robust tool for aiding clinicians in the early and accurate diagnosis of AD.

Main Methods:

• A convolutional neural network (CNN) with a U-Net architecture was trained using the nnU-Net framework on a dataset of 70 confirmed AD and 87 non-AD cases.
• The model was validated on an internal test dataset of 106 cases and an external dataset comprising 100 AD and 38 non-AD cases.
• Performance was evaluated using AUROC, AUPRC, sensitivity, specificity, precision, and F1-score.

Main Results:

• The AI algorithm achieved high performance, with an AUROC of 98.7% on the internal test set and 97.0% on the external test set.
• On the external test dataset, the algorithm demonstrated 92.0% sensitivity, 100.0% specificity, 100.0% precision, and a 95.8% F1-score.
• The AI successfully detected 93.3% of unsuspected AD cases in the internal test set.

Conclusions:

• Optimized CNNs can reliably detect AD across diverse, multicenter CT imaging datasets.
• The developed AI pipeline demonstrates the potential for streamlined, robust AD detection on CT scans.
• The AI tool will be publicly available for further research and clinical evaluation.