Multi-Stage Cascaded Deep Learning-Based Model for Acute Aortic Syndrome Detection: A Multisite Validation Study.

Area of Science:

• Radiology
• Artificial Intelligence
• Cardiovascular Imaging

Background:

• Acute Aortic Syndrome (AAS), comprising aortic dissection (AD), intramural hematoma (IMH), and penetrating atherosclerotic ulcer (PAU), poses diagnostic challenges.
• Rapid assessment is critical for managing AAS due to its varied presentations.

Purpose of the Study:

• To develop and validate a multi-stage deep learning model for detecting AAS on chest computed tomography angiography (CTA) scans.
• To assess the model's performance across diverse patient demographics and imaging conditions.

Main Methods:

• A U-Net architecture was used for aortic segmentation, followed by cascaded classification for AD/IMH and a multiscale CNN for PAU detection.
• External validation involved 260 anonymized CTA scans from 14 U.S. clinical sites, covering four CT manufacturers.
• Performance was evaluated using sensitivity, specificity, and AUC, with 95% confidence intervals calculated via Wilson's method.

Main Results:

• The model achieved high performance for overall AAS detection: sensitivity 0.94, specificity 0.93, and AUC 0.96 (all p < 0.001 vs. 0.80 benchmark).
• Consistent performance was observed across subgroups, including different patient demographics, CT manufacturers, slice thicknesses, and anatomical locations.
• The model demonstrated robust detection of AD, IMH, and PAU.

Conclusions:

• The developed deep learning model effectively detects the full spectrum of AAS.
• The model's consistent performance across diverse populations and imaging platforms suggests significant clinical utility.
• This AI tool has the potential to expedite patient triage and management in clinical settings.