Deep learning detected cerebral microbleeds accurately, but in only five patient-level studies

Design

Systematic review and diagnostic accuracy meta-analysis

Population

Studies of MRI-based deep-learning models for cerebral microbleeds (5 patient-level)

Primary outcome

Sensitivity and specificity

Effect

Patient level: sensitivity 0.89 (0.76 to 0.96), specificity 0.86 (0.77 to 0.92)

This meta-analysis pooled studies of deep-learning models detecting cerebral microbleeds on MRI. At patient level, five studies gave sensitivity 0.89 (95% CI 0.76 to 0.96) and specificity 0.86 (0.77 to 0.92). At lesion level, sensitivity was 0.96 and specificity 0.98.

Microbleed counting matters more than it used to: it informs anticoagulation decisions, and it is required before and during anti-amyloid treatment for Alzheimer disease, where ARIA monitoring depends on it. Manual counting is slow and variable.

The numbers look good, but five patient-level studies is a small base, and lesion-level metrics can flatter a model because most of the brain is easy to call negative. External validation on local scanners is essential before clinical use.

• Microbleed counts on susceptibility-weighted imaging inform anticoagulation and anti-amyloid decisions
• An AI tool can assist counting but a radiologist should confirm borderline lesions
• Check any tool against your own scanners and sequences before relying on it
• Patient-level sensitivity of 0.89 means about one in ten affected patients may be missed

Why it matters

Anti-amyloid therapy is turning microbleed counting into a high-volume task.

Don't overread it

Only five patient-level studies; the lesion-level accuracy overstates what a clinician would experience.

The statistics, in plain English

Sensitivity 0.89 means 89 of 100 patients with microbleeds are flagged; specificity 0.86 means 14 of 100 without them are wrongly flagged. The wide intervals reflect the small number of studies.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.