Related Experiment Video
Updated: May 3, 2026
23:33
The WATCHMAN Left Atrial Appendage Closure Device for Atrial Fibrillation
Published on: February 28, 2012
83.3K
Detecting the left atrial appendage in CT localizers using deep learning
Aydin Demircioğlu1, Denise Bos2, Anton S Quinsten2
1Institute of Diagnostic and Interventional Radiology and Neuroradiology, University Hospital Essen, Hufelandstraße 55, 45147, Essen, Germany. aydin.demircioglu@uk-essen.de.
Scientific Reports
|May 2, 2025
Summary
Deep learning accurately identifies the left atrial appendage (LAA) in CT localizer scans, improving accuracy and reducing radiation exposure. This automated method enhances patient safety by minimizing unnecessary whole-heart scans.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Radiology
- Cardiovascular Imaging
Background:
- Cardioembolic stroke patients often require CT scans of the left atrial appendage (LAA) to detect thrombi.
- Current LAA imaging relies on localizer scans, but imprecise landmark identification necessitates whole-heart scans, increasing radiation exposure.
- Accurate LAA delimitation in localizers is crucial for targeted imaging and radiation dose reduction.
Purpose of the Study:
- To develop and evaluate a deep learning model for automated left atrial appendage (LAA) delimitation in CT localizer scans.
- To compare the accuracy and clinical utility of different deep learning networks for LAA boundary prediction.
- To assess the potential for radiation dose reduction using AI-guided LAA imaging.
Main Methods:
- Retrospective collection of 1253 CT localizer scans from a single center.
- Training and evaluation of four deep learning networks: VariFocalNet, Cascade-R-CNN, Task-aligned One-stage Object Detection Network, and YOLO v11.
- Validation on internal (368 patients) and external (309 patients) test cohorts, assessing accuracy (Dice coefficient) and clinical utility.
Main Results:
- VariFocalNet demonstrated superior performance, achieving high LAA delimitation accuracy (97.8% internal, 96.8% external) and Dice coefficients (90.4% internal, 90.0% external).
- Clinical utility was near-perfect (99.8% internal, 99.3% external).
- Network-based delimitation reduced radiation exposure by over 50% compared to whole-heart scanning in both internal and external cohorts.
Conclusions:
- Deep learning models, particularly VariFocalNet, can accurately and reliably delimit the left atrial appendage (LAA) in CT localizers.
- Automated LAA delimitation significantly enhances imaging accuracy and reduces patient radiation exposure.
- This AI-driven approach offers a promising alternative to conventional whole-heart scanning for LAA assessment in stroke patients.
