SAM-guided multi-view network automates intracranial aneurysm segmentation on 3D TOF-MRA with multicenter validation

A SAM-guided deep learning model segmented intracranial aneurysms on 3D TOF-MRA with external-test sensitivity 83.4% (95% CI 75.9–88.0) and Dice 0.84 (95% CI 0.79–0.87), outperforming junior radiologists and matching mid-level readers.

• Internal development set comprised 1,560 patients (1,554 aneurysms); internal test 421 patients (522 aneurysms); independent external test set 208 patients (224 aneurysms) with DSA as reference standard.
• Internal test performance: lesion-level sensitivity 86.0% (95% CI 82.8–88.7), precision 87.4% (95% CI 84.2–90.0), Dice similarity coefficient 0.88 (95% CI 0.84–0.90).
• Overall detection F1-score 86.7; significantly outperformed junior radiologists, statistically comparable to mid-level radiologists, inferior to senior radiologists; average processing time about 72 seconds per exam.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.