DL model reconstructs complete cerebral arteries from non-contrast CT for pre-thrombectomy planning

A deep-learning model (nnU-Net) reconstructed the full cerebral vasculature from non-contrast CT in large-vessel-occlusion stroke patients, achieving a Dice score of 0.79±0.04 on external validation. Radiologists rated 98.9% of segmentations as high-quality. Prospective studies…

• Retrospective training on 280 paired NCCT-CTA exams without occlusion; externally validated on 40 exams, then tested against DSA in 290 LVO-AIS patients from two hospitals.
• Mean Dice coefficient was 0.80±0.04 (internal validation), 0.79±0.04 (internal test), and 0.79±0.04 (external validation).
• Study is retrospective and lacks prospective validation; clinical feasibility is shown but impact on thrombectomy outcomes is unproven.

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