The labeling gap, in one dataset
In December 2024, researchers from Boston Children's Hospital published a review of all 876 AI/machine-learning-enabled medical devices the FDA had authorized through early 2024, in JAMA Pediatrics. Only 149 devices (17.0%) were labeled for pediatric use. Roughly a third (33.3%) were authorized for adults only, and about half (49.7%) carried a "silent" label that said nothing about pediatric use either way.
Radiology dominates the devices that do carry a pediatric label — 119 of the 149 (79.9%), ahead of neurology (8.7%) and cardiovascular (6.7%). That tracks with radiology's share of AI/ML devices overall (76.1% of all 876). So the gap isn't a niche specialty problem; it sits squarely in imaging.
The labeling gap also understates a validation gap. Among the 149 devices labeled for pediatric use, only 28 (18.8%) were actually validated on pediatric data. Another 22 (14.8%) were validated using adult data only, and for the remaining 99 (66.4%) the manufacturer didn't report whether pediatric patients were included in validation at all. A pediatric label, in other words, is not reliable proof that a model was ever tested on a child's anatomy.
What happens when an adult-trained model reads a child's scan
A scoping review published in European Radiology in February 2026 pulled together 20 studies testing adult-trained AI models — unmodified — against 16 pediatric imaging datasets ranging from 30 to 7,357 subjects, across radiography, CT, MRI, DEXA and ultrasound. Apart from two studies, every one reported a performance drop when the adult model was applied to children.
The degradation wasn't uniform, but it was consistent in direction. One deep-learning pulmonary nodule detector's sensitivity fell from 95.5% in adults to 53.1% in children. Lung-boundary segmentation accuracy dropped from 95.4% in adults to 36.8% in children under 2. Across organs and modalities, segmentation Dice scores fell by more than 0.10, and in children aged 0–2 some Dice scores ran as low as 0.07–0.65, against 0.80–0.92 in adults. A pneumonia classifier's AUC fell from 0.95 in adults to 0.82 in children. The pattern the reviewers flagged most: children 2 years old and younger consistently showed the largest deficits across every task type — detection, segmentation and classification alike.
None of this means adult-trained models are unusable for pediatric imaging. It means they can't be assumed to generalize, and the review's authors point to specific fixes that do measurably help: fine-tuning on pediatric-only data (reported Dice-score gains of 47–79%), adjusting detection thresholds by age and lesion type (one pneumothorax detector's sensitivity in children under 2 rose from 50% to 60% after retuning — a real gain, though still short of adult-level performance), and narrowing a tool's scope to exclude the youngest patients and highest-risk misses, which in one series lifted overall accuracy from 87.5% to 96.9%.
A domain-adaptation method tested across 30 NICUs in 20 countries
The newest entry in this line of research, published in Nature Biomedical Engineering on October 13, 2026 by Vishwanatha Rao, Pranav Rajpurkar and a multi-institution team based at Harvard Medical School, targets a specific obstacle: pediatric imaging data is scarce and hospitals are reasonably reluctant to pool it with outside institutions.
Their method — continual learning — passed a single deep-learning model, originally trained to assess endotracheal tube placement on adult chest X-rays, sequentially through 30 neonatal intensive care units in 20 countries and territories. Each site retrained the model on just 50 locally held neonatal chest radiographs (2,965 images in total) inside its own environment, then passed the updated model to the next hospital — no patient image ever left the building where it was taken.
The result: average tube-placement error fell to 0.67 cm for the continually adapted model, versus 1.29 cm for a model fine-tuned at a single hospital alone and 4.30 cm for the original, unadapted adult model — an error margin large enough to miss a dangerously misplaced tube in a newborn. The adapted model beat the single-site approach at 29 of the 30 test sites, with improved detection of dangerously deep intubations at 27 of 30. The researchers frame the work against the same backdrop: fewer than one in five commercially available models are cleared for paediatric use — the same gap the JAMA Pediatrics data quantified two years earlier.
Comparing the adaptation approaches
| Method | What it does | Reported effect |
|---|---|---|
| Fine-tuning on pediatric data | Retrains an adult model on a pediatric-only dataset | Dice-score gains of 47–79% (European Radiology, 2026) |
| Operating-point adjustment | Recalibrates detection thresholds by age and lesion type | Pneumothorax sensitivity in children ≤2y: 50% → 60% |
| Selective scoping | Excludes youngest patients or highest-risk findings from automated reads | Accuracy rose from 87.5% to 96.9% in one pediatric chest-radiograph series |
| Continual learning (multi-site) | Adapts one model sequentially across hospitals; data stays local | Tube-placement error: 4.30 cm → 0.67 cm (Nature Biomedical Engineering, 2026) |
The buyer-side version of this problem
Most of the coverage of the pediatric AI gap is framed as a research or regulatory story. For an outpatient imaging center or community hospital, it's a procurement story. Dedicated children's hospitals read a small share of U.S. pediatric imaging volume — most pediatric chest and abdominal CT, and nearly all pediatric emergency imaging, is read in general radiology departments and outpatient centers that also serve adults. Those are exactly the sites an AI-reporting vendor is least likely to have validated for.
That creates a specific due-diligence question that most vendor evaluation checklists don't ask yet: not just "is this FDA-cleared," but "cleared and validated for which ages, and what happens to a pediatric study if the model wasn't trained for it." Given that two-thirds of pediatric-labeled devices don't report their pediatric validation status at all, "pediatric-labeled" alone isn't a sufficient answer — buyers reading any pediatric volume should ask the vendor directly, in writing, rather than infer it from the FDA label.
Three questions follow directly from the data above, for any site that reads even occasional pediatric studies:
Was the model validated on pediatric data — specifically, which ages?
Performance deficits concentrate in children 2 and under. A vendor's adult-wide accuracy numbers say nothing about performance in that age band unless pediatric cases, broken out by age, were part of validation.
How are pediatric studies flagged or routed?
If a model was not built or validated for pediatric imaging, the safer design routes pediatric studies to a different workflow or additional review rather than silently applying an adult-calibrated model.
Is there a documented adaptation or generalizability plan?
The research above shows pediatric adaptation is solvable with fine-tuning, threshold recalibration, or multi-site continual learning. A vendor without any pediatric-specific plan is applying an adult-only tool to a population it was not built for.
Where xAID fits
xAID's CT-reporting AI is built around the same principle the research above keeps landing on: a model's reported accuracy is only meaningful for the population it was actually built and tested on, and every report needs a qualified radiologist to catch what a model gets wrong. In practice that means xAID's in-house radiologist reviews every preliminary report before it reaches a client ready-to-sign — the same human-in-the-loop layer that a pediatric-unvalidated model needs most, not least.
Frequently asked questions
What percentage of FDA-authorized AI imaging devices are labeled for pediatric use?
About 17%. A 2024 JAMA Pediatrics review of 876 FDA-authorized AI/ML-enabled medical devices found that only 149 (17.0%) were labeled for pediatric use, while a third were authorized for adults only and about half made no reference to pediatric use at all.
Does adult-trained radiology AI perform as well on children as on adults?
No. A 2026 European Radiology scoping review of 20 studies found that adult-trained AI models lost accuracy on pediatric imaging in nearly every case, with the steepest drops in children under 2 years old — for example, pulmonary nodule detection sensitivity falling from roughly 95% in adults to about 53% in children in one study.
What is the new domain-adaptation method for pediatric imaging AI?
Researchers at Harvard Medical School, publishing in Nature Biomedical Engineering in October 2026, used continual learning to adapt an adult endotracheal-tube-placement model for neonatal chest X-rays across 30 NICUs in 20 countries, with each hospital retraining the model locally so patient images never left the building. The adapted model's placement error fell to 0.67 cm, versus 1.29 cm for single-hospital fine-tuning and 4.30 cm for the unadapted adult model.
What does the pediatric AI labeling gap mean for outpatient imaging centers and community hospitals?
Most pediatric CT and chest-radiograph volume in the U.S. is read outside dedicated children's hospitals, in outpatient centers and community facilities whose AI-reporting tools are overwhelmingly validated on adults. Buyers evaluating an AI-reporting vendor for a site that reads any pediatric studies should ask specifically whether the model was validated on pediatric data, at what ages performance was tested, and how pediatric cases are flagged or routed if the vendor cannot answer that.
Source: Brewster et al., "US FDA Approval of Pediatric Artificial Intelligence and Machine Learning–Enabled Medical Devices," JAMA Pediatrics (2024), doi.org/10.1001/jamapediatrics.2024.5437. Laborie et al., "Performance of adult-trained artificial intelligence models in paediatric imaging—a scoping review," European Radiology (2026), doi.org/10.1007/s00330-026-12354-5. Rao, Rajpurkar et al., "Repurposing adult AI radiograph models for neonatal care through continual learning: an international multicentre study," Nature Biomedical Engineering (2026), doi.org/10.1038/s41551-026-01812-6, as covered by Radiology Business. Figures are rounded as reported.