In a retrospective diagnostic accuracy study of 485 adults with suspected acute lacunar stroke, a deep learning model applied to routine CT perfusion beat stroke-neurologist consensus on the held-out validation set (AUC 0.82 vs mean clinician AUC 0.58; p=0.003). The study, conducted across 2 comprehensive stroke centers in New South Wales, used diffusion-weighted MRI as the reference standard for clinically relevant lacunar infarct.

Why It Matters To Your Practice

• Lacunar stroke can be hard to confirm acutely when MRI is unavailable, delayed, or impractical.

• CT perfusion is widely available, but this study found poor clinician performance and low interobserver agreement for lacunar-stroke detection on CTP (Fleiss κ=0.22).

• An automated model that improves consistency on routine imaging could help triage uncertain cases in acute Cerebrovascular accident (Stroke) workflows.

Clinical Implications

• The best-performing model used cerebral blood flow maps alone, suggesting AI support may not require complex multiparametric interpretation.

• This tool is best viewed as decision support, not a replacement for MRI, bedside exam, or overall stroke assessment.

• Potential use cases include flagging likely lacunar infarcts, supporting transfer/imaging decisions, and reducing variability among readers.

Insights

• Of 485 eligible patients, 436 were used for model development and 49 formed the held-out validation set.

• The validation cohort included 27 lacunar strokes and 22 mimics, with diffusion-weighted MRI defining the reference diagnosis.

• The authors concluded the model's diagnostic accuracy was comparable to or better than expert stroke neurologists, but called for prospective and external validation before clinical deployment.

The Bottom Line

• For suspected acute lacunar stroke, deep learning on routine CTP outperformed neurologist consensus in this small held-out sample.

• Clinicians should see this as a promising reliability tool for acute stroke practice — especially where MRI access is limited — not as practice-changing evidence yet.