Deep Learning Model for 3D Segmentation of Vertebrobasilar Artery Calcification on CBCT Achieves Mean Dice 0.73

ImprovedVertebroV5, a 3D U-Net optimized for small structures, segmented vertebrobasilar artery calcification on cone-beam CT with mean Dice 0.73 and small-object sensitivity 0.94 in 20 patients (retrospective, single-center). Significant gains over prior model.

• Retrospective, internal validation on 20 patients undergoing vertebrobasilar evaluation with CBCT; training set size not reported.
• High small-object sensitivity (0.94 ± 0.18) indicates accurate detection of tiny calcifications; precision 0.78, recall 0.75, specificity 0.9997.
• No external validation; single-center study with small test set limits generalizability; manual reference standard may introduce variability.

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