Members of our team have contributed to a significant academic study Titled "Automated LVO Detection and Collateral Scoring on CTA Using a 3D Self-Configuring Object Detection Network: A Multi-Center Study," this research employed the nnDetection model, a state-of-the-art self-configuring object detection network. This model was rigorously trained on single-phase CTA scans from 2,425 patients across five centers and subsequently validated on an external test set of 345 patients from an additional center.

The results are highly promising: the nnDetection model achieved an impressive diagnostic accuracy of 98.26% (95% CI 96.25–99.36%) in identifying LVO. Furthermore, it correctly classified 339 out of 345 CTA scans in the external test set. The deep learning-based collateral scores showed a kappa of 0.80, reflecting good agreement with the evaluations provided by three expert radiologists.

These findings demonstrate that the nnDetection model can not only detect LVO with high accuracy on single-phase CTA scans but also provide semi-quantitative assessments of collateral flow, offering a comprehensive solution for automated stroke diagnostics in patients with LVO. This advancement holds great promise for improving the speed and accuracy of stroke diagnosis, ultimately enhancing patient outcomes.

 For details: https://www.nature.com/articles/s41598-023-33723-w