From late July to early November 2026, three papers on obstetric ultrasound AI were published online in succession. Their conclusions sound similar—“AI helps”—yet their study designs differ by orders of magnitude: one is a randomized crossover trial in China using 1,584 real scans; another is an image-interpretation trial involving 96 Danish clinicians (and incidentally reveals a cautionary insight about “humans”); and the third tests its model on only 17 patients—and reports a “perfect” 100% sensitivity. Laid side by side, these three reports collectively form a ready-made teaching specimen of the evidence hierarchy in medical AI—and answer a practical question: How trustworthy is fetal ultrasound AI *right now*?
## First Report Card: A Randomized Trial in Real Clinical Settings
The PAICS (Prenatal Ultrasound Diagnosis Artificial Intelligence Conduct System) trial, published online in *The Lancet Digital Health* on September 10, was conducted across five centers in China. Participants were singleton pregnancies at high risk for fetal anomalies, gestational age 11–32 weeks. The sonographers performing the scans had 3–8 years of clinical experience. The trial’s greatest strength lies in its design: a multicenter, self-crossover design—each participant was randomly assigned 1:1 to one of two diagnostic sequences: either “independent real-time scanning followed by AI-assisted offline review,” or vice versa, with a four-week washout period between sequences so that each sonographer served as their own control under both conditions. Allocation was concealed; sonographers were blinded to fetal anomaly status; and outcome assessors—as well as the independent expert panel (who reviewed video clips) serving as the reference standard—were blinded to group assignment. The primary endpoint was deliberately “obstetric”: sensitivity and specificity for detecting ten specific fetal intracranial anomalies—sensitivity was tested for superiority, while specificity was assessed against a non-inferiority margin of 5%. Both missed diagnoses and false positives—the two distinct costs—were explicitly addressed at the design stage.
The results are clean: 1,584 scans were completed between September 6, 2022, and November 1, 2023. In the fetus-level analysis, AI assistance increased sensitivity by 0.087 (95% CI 0.029–0.147; P<0.0001 for superiority), while specificity changed by only 0.009 (CI −0.006 to 0.023), meeting the non-inferiority criterion. In the anomaly-targeted analysis, sensitivity increased by 0.118 (CI 0.054–0.181), with specificity again non-inferior. No adverse events related to the diagnostic process or AI assistance occurred. The authors’ conclusion remains restrained: PAICS improves detection sensitivity for targeted intracranial anomalies without compromising specificity, supporting integration of AI assistance into fetal anomaly screening for high-risk scenarios. The study was funded by the National Natural Science Foundation of China and grants from Guangdong Province, Guangzhou City, and Sun Yat-sen University; authors declared no competing interests. One author’s affiliation is listed as an information technology company—a declaration and an affiliation are two independently traceable pieces of information, worth reading separately.
But boundaries must be seen clearly: anomalies were limited to the ten intracranial targets identifiable by the AI; the population was restricted to high-risk singleton pregnancies across five Chinese centers; and sonographers had 3–8 years of experience—extrapolation to novices or senior experts remains unvalidated.
## Second Report Card: An Image-Interpretation Trial—and a Warning About “Humans”
An AJOG MFM paper (published online November 6) reports a trial conducted across five hospitals in Denmark: 96 obstetric and gynecologic clinicians (residents, gynecologists, and obstetricians) were randomized 1:1 to receive or not receive AI-generated visual and textual feedback while interpreting standardized umbilical artery and middle cerebral artery Doppler images. Each clinician interpreted 40 standardized images; the 96 participants were paired into 48 matched pairs based on obstetric ultrasound experience, yielding a total of 1,920 unique image interpretations. The interpretation task merits clarification: for color Doppler images, participants marked where the gate should be placed; for spectral Doppler images, they judged whether the waveform was technically adequate—this assessed “how well this measurement was performed,” not fetal diagnosis. Reference answers were derived from expert consensus: two fetal medicine sonographers with >10 years’ experience applied pre-specified standards from the International Society of Ultrasound in Obstetrics and Gynecology (ISUOG); disagreements were adjudicated by a fetal medicine consultant. AI feedback itself was classified as “correct” or “incorrect” depending on whether it aligned with expert consensus. “Novices” were defined as having <1 year of obstetric ultrasound experience; “experienced clinicians” had ≥1 year.
Results: AI feedback raised overall interpretation accuracy from 62.3% to 75.0% (difference 12.7 percentage points; 95% CI 8.2–17.1; P<0.001). Novices (n=36) improved by 12.1 percentage points (70.0% vs. 57.9%), and experienced clinicians (n=60) improved by 13.0 percentage points (78.0% vs. 65.0%)—both groups benefited. Yet three details deserve closer attention. First, AI alone outperformed novices without feedback. Second, experienced clinicians + AI outperformed AI alone—clinical experience was not replaced by the algorithm. Third, when AI feedback was incorrect, novices followed the error significantly more often than experienced clinicians (22.3% vs. 15.8%; P=0.01)—automation bias was deeper among those least experienced. Accordingly, the authors’ conclusion carries explicit caveats: clinical experience remains critical when AI feedback is wrong; human–AI collaboration requires further evaluation in real-time Doppler acquisition settings. Note also that this trial assessed standardized image interpretation—not bedside real-time scanning. The gap from exam paper to patient bed remains unvalidated.
## Third Report Card: A “Perfect” Score on Just 17 People
A study on placenta accreta spectrum (PAS) published in the October issue of *AJOG MFM* (online July 31) addresses a real and life-threatening problem: PAS is a leading cause of maternal morbidity and mortality, its incidence is rising, and prenatal diagnostic accuracy remains only 30%–50%. The team extracted 38,907 grayscale frames from 756 DICOM files of 113 PAS-risk pregnant women enrolled between 2018 and 2025 at a single center (mean gestational age at ultrasound: 30.89±3.67 weeks), labeled each frame according to final pathological grading, and split patients into training (n=79), validation (n=17), and test (n=17) sets. Using pathology as the gold standard carries an implicit cost: only cases reaching pathological confirmation entered the dataset—selection bias cannot be ruled out. The model was not purely image-based: each frame was first scored probabilistically using an ImageNet-pretrained EfficientNetB0; frame-level scores were aggregated into patient-level scores, then combined with two clinical variables—prior cesarean deliveries and placenta previa status—and fed into an ensemble classifier composed of logistic regression, random forest, and gradient boosting.
Test-set performance looks flawless: accuracy 88%, sensitivity 100%, negative predictive value 100%, zero false negatives, AUC 0.972. Model feature importance centered on the placental interface—an observation the authors cite as biological plausibility. But displaying confidence intervals changes the story: the 95% CI for sensitivity is 66.4%–100.0%; for specificity (75%), it is 34.9%–96.8%—with only 17 people in the test set, any “perfect” score has a confidence interval wide enough to drive a truck through. The authors themselves position this as a proof-of-concept: results support its potential as an earlier screening tool, but prospective trials are needed. The correct way to read this study is in one sentence—“promising, not proven.”
## My Three Judgments
Together, these three report cards yield three judgments—synthesized from integrating all three bodies of evidence, not drawn from any single paper’s conclusions.
First, the evidence hierarchy is not rhetorical—it reflects differences in scale. All three claim “obstetric ultrasound AI is effective,” yet one is a randomized, self-crossover trial with 1,584 real scans, another is an image-interpretation trial with 96 participants, and the third is a retrospective test set of just 17 patients. Their credibility gaps dwarf their superficial similarity in titles. Readers and decision-makers should weigh evidence by study design—not by journal name or headline accuracy figures. And one blank common to all three studies remains glaring: none measured patient outcomes—no study verified whether improved detection sensitivity actually altered pregnancy outcomes; “no adverse events” does not equal long-term safety evidence.
Second, AI raises the floor—but charges the highest toll on those least experienced. The Danish trial provides a complete ability ranking: experienced clinicians + AI > AI alone > novices without assistance; and when AI erred, novices followed the error ~40% more often than experienced clinicians (22.3% vs. 15.8%). The direct implication for training is clear: before introducing AI assistance, train novices first on “when and how to disagree with AI”—otherwise the tool risks deepening automation bias precisely among those most vulnerable to it.
Third, obstetric screening deserves its own evidence template: superiority for sensitivity *plus* non-inferiority for specificity. Missing a catastrophic anomaly carries devastating consequences; false positives trigger anxiety and unnecessary follow-up—these costs are asymmetric. PAICS’s dual-threshold design directly confronts this asymmetry and deserves adoption as the default template for obstetric AI trials; the same yardstick also exposes the true substance of the “perfect” 100% sensitivity on 17 people—its confidence interval is too wide to sustain the word “perfect.”
## What You Can Take Away
If you are a sonographer or obstetrician: AI assistance for intracranial anomaly screening in high-risk pregnancies now has randomized trial evidence grounded in real scanning; however, AI feedback for Doppler assessment remains at the image-interpretation level only—and if your experience is <1 year, be especially wary of the impulse to “follow the AI.” In the trial, this group followed erroneous feedback most frequently.
If you are a hospital administrator: Prioritize deployment in high-risk screening contexts; embed “questioning AI” into training and evaluation—not just “using AI.”
If you are a researcher: Self-crossover design is worth emulating; for 100% sensitivity reported on a 17-patient test set, always ask first for the confidence interval.
If you are an expectant parent: AI assistance affects only the detection sensitivity step; reports remain signed and clinically accountable to physicians. What matters more on your prenatal report is whether all expected anatomical structures were clearly visualized—not whether AI was used during the scan.
The truly debatable question is: When AI assistance lifts anomaly detection rates, who should define the subsequent care pathway—follow-up, re-evaluation, and intervention decisions? Physicians? Hospitals? Or clinical guidelines? Join the discussion in the comments.
Source: Multicenter, Self-Crossover, Randomized Controlled Trial of AI-Assisted Detection of Fetal Intracranial Anomalies (Lancet Digit Health 2026;8(10):101040). https://doi.org/10.1016/j.landig.2026.101040