Deep reinforcement learning for automatic anatomic CT landmark localization in Stanford Type B aortic dissection.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Cardiovascular imaging

Background:

• Consistent, landmark-based measurements are crucial for long-term monitoring of aortic dissection.
• Stanford Type B aortic dissection (TBAD) requires precise anatomical assessments over time.

Purpose of the Study:

• To evaluate deep reinforcement learning (DRL) agents for detecting anatomical landmarks in patients with TBAD.
• To compare the performance of DRL agents against human interobserver variability.

Main Methods:

• Retrospective analysis of 396 CT angiography scans from 9 international sites.
• Manual labeling of aortic landmarks (annulus, 8 branch vessels) and interobserver variability assessment.
• Training DRL agents for landmark detection, comparing single-agent and cluster-based prediction algorithms.

Main Results:

• DRL single agents achieved a median error of 2.7 mm with a 4.8% failure rate.
• Cluster-based DRL models demonstrated superior performance with a median error of 2.5 mm and 4.0% failure rate.
• Cluster-based DRL predictions were significantly better than single-agent predictions and comparable to human interobserver variability.

Conclusions:

• DRL agents, particularly cluster-based models, accurately and precisely predict aortic landmarks in TBAD patients.
• The performance of DRL models rivals human observer variability, offering a potential tool for efficient aortic dissection monitoring.
• DRL provides a rapid (median 1.0 second processing time) and reliable method for landmark detection in TBAD imaging.