AI assistance raised fetal-brain abnormality detection in a 1,584-scan trial

Sonographers detected more targeted fetal brain abnormalities with AI assistance in a randomised trial involving 1,584 scans at five Chinese centres. The PAICS system improved fetal-based sensitivity by 8.7 percentage points while preserving specificity within the trial’s prespecified margin.[1][2]

Twenty-nine sonographers, each with between three and fewer than eight years of experience, assessed pregnancies at high risk of fetal malformation. The study, published in The Lancet Digital Health in September 2026, compared performance with and without PAICS, a system designed to help identify ten specified intracranial abnormalities. For the fetal-based analysis, sensitivity rose from 78.6% without AI to 87.3% with it—an 8.7-percentage-point gain.[1][2]

The distinction between finding more and alarming everyone

Sensitivity answers how often a diagnostic process identifies a condition among cases in which it truly exists. Specificity asks how often it correctly rules the condition out when absent. Raising one measure at the expense of the other can create a flood of false alarms. The researchers therefore established two requirements in advance: superiority in sensitivity and non-inferiority in specificity within a five-percentage-point margin.[1][2]

Both primary requirements were met. Fetal-based specificity was 96.0% without assistance and 96.9% with it. The sensitivity difference had a reported 95% confidence interval of 2.9–14.7 percentage points. A separate, malformation-targeted analysis found an 11.8-point sensitivity improvement, from 69.6% to 81.4%; its unit of analysis differs from the fetal-based result.[1][2]

It was the clinician and the model together

The experiment used a randomised self-crossover sequence. Each case received both assisted and unassisted diagnostic assessment, in an ordering determined by randomisation, and an expert panel's video review served as the reference standard. Sonographers were aware when they had AI support; experts assessing the reference diagnosis were masked to assignment. The structure matters because the research was evaluating a combined workflow, not a machine diagnosing pregnancies entirely on its own.[1][2]

Indeed, clinicians sometimes corrected the machine. The publication reports that when PAICS generated false positives or false negatives, sonographers overrode a substantial proportion: 53 of 83 false positives and 38 of 64 false negatives. That is a different, more credible conception of assistance from an uncritical 'AI decides' workflow. The model can draw attention to a structure; an experienced operator can use surrounding information and judgement to accept or reject the flag.[1][2]

Who was studied, and who was not

The study enrolled high-risk singleton pregnancies scanned at 11–32 weeks and sonographers with three to fewer than eight years of experience. Its ten-condition target and defined operator group matter when applying the result to a different screening service.[1][2]

The measured outcome was detection during scanning. The trial did not follow whether the additional findings changed pregnancy or newborn outcomes. It reported no adverse events related to the diagnostic process or AI assistance.[1][2]

A reproducible gain in detection with preserved specificity is a useful clinical result. Further deployment studies can now ask how well it carries into other equipment, staffing and case mixes.

Report an error or suggest a correction