The warning: a well-validated criteria set, applied to the wrong population
On October 14, 2026, the American College of Radiology published guidance, authored by Stamatia Destounis, MD, FACR, and colleagues, on what it called the misuse of BI-RADS assessment categories on pediatric and adolescent breast masses. Breast lesions are not rare in children — the guidance puts their incidence at roughly 0.1% to 3.25% of the pediatric and adolescent population — but primary breast malignancy in that age group is exceedingly rare, with an incidence of about 0.08 newly diagnosed cases per 100,000 people.
The ACR's point is narrow and specific: BI-RADS ultrasound descriptors can still be useful for describing what a pediatric mass looks like, but the BI-RADS assessment categories — and the malignancy probability each one implies — were developed and validated for screening adult populations, not for diagnostic workup of a palpable mass in a child. As Destounis put it: "Children and adolescents are not simply small adults. Using adult risk assessment tools in young patients may create unnecessary concern and lead to procedures that do not improve care."
What happens when the category doesn't fit the population
The ACR guidance points to a concrete illustration: a multicenter study of patients younger than 21 who underwent pre-operative breast ultrasound, published in the Journal of Pediatric Surgery in 2021. Of 227 patients, 84% were assigned a BI-RADS category, and BI-RADS 4 — the category that, by adult definition, carries a 2% to 95% likelihood of malignancy — was the single most frequently assigned category, given to 55% of the cohort. On pathology, every lesion in the study was benign. The investigators, writing in their paper, concluded that adult BI-RADS categories and their corresponding biopsy recommendations should not be routinely applied to pediatric patients.
That gap is the mechanism, not an anomaly. BI-RADS 4's malignancy range was calibrated against adult breast-cancer prevalence. Drop a population with a baseline cancer risk near zero into that same categorical scale, and the scale stops meaning what it was built to mean — a mass that would sit comfortably in a low-concern bucket in an adult-calibrated world gets pushed into a "probably biopsy this" category by the structure of the criteria itself, not by the biology of the lesion.
The cost of a mismatched category isn't abstract
The ACR guidance is explicit about the downstream effects of applying adult-calibrated BI-RADS 4 thresholds to pediatric lesions: unnecessary biopsy or excision with potential injury to developing breast tissue, recommendations for additional surveillance imaging that consume resources and drive up cost, and patient and family anxiety that a near-zero-risk finding does not warrant. None of that reflects a failure of ultrasound as an imaging modality — the ACR still recommends targeted ultrasound as the preferred pediatric exam because it avoids ionizing radiation. The failure sits specifically in using an adult-validated category-to-risk mapping as if it transfers unchanged to a population it was never tested on.
Adult-validated criteria, applied as-is vs. adapted to the population
| Dimension | Applied as-is to a new subpopulation | Adapted to the subpopulation |
|---|---|---|
| Baseline risk assumption | Inherited from the original (adult) validation cohort | Re-grounded in the actual population being read |
| Category-to-risk mapping | Unchanged — same thresholds, new population | Recalibrated or supplemented with population-specific context |
| Observed outcome (pediatric BI-RADS 4 example) | 55% of a 227-patient cohort flagged as "probably malignant" range | 0% of that same cohort was actually malignant |
| Downstream effect | Unnecessary biopsy, surveillance imaging, cost, anxiety | Management guided by clinical presentation and population-specific evidence |
Why this is a structured-reporting problem, not just a breast-imaging one
BI-RADS is one of several structured reporting-and-data systems radiology relies on — alongside analogues like Lung-RADS, PI-RADS, TI-RADS, and O-RADS — that compress a radiologist's impression into a standardized category with an implied management pathway. That standardization is valuable precisely because it's consistent: a BI-RADS 4 means roughly the same thing from one adult mammography report to the next. The pediatric case shows what breaks when that consistency gets exported to a population the schema was never validated against: the category stays the same, but what it predicts does not.
That is a design problem any structured-reporting system has to account for, whether the categories are assigned by a radiologist applying a published atlas or generated by software. A risk category, a severity bucket, or a recommended-follow-up template is only as trustworthy as the population it was calibrated on. Age is one axis where that calibration can fail, as the ACR guidance shows — but anatomy, scanner protocol, and clinical indication can just as easily shift a population away from whatever baseline a reporting template assumed.
Where this applies to AI-generated reports
The same risk applies directly to AI systems that generate structured report drafts: a model trained and tuned on an adult chest CT population will carry those population assumptions into its findings and risk language unless it's explicitly built to recognize when a study falls outside that calibration — different age group, different anatomy, different clinical context — and adapt accordingly, rather than forcing every case through one static template. A foundation-model approach that reasons across a full study, instead of applying a narrow, rigidly-templated detector built for one population, is better positioned to flag that mismatch rather than silently misapply it. It's also exactly the kind of failure mode that keeps a radiologist in the review loop: xAID's reports are generated as structured drafts and reviewed by xAID's in-house radiologist before reaching a client ready-to-sign, so a category that doesn't fit the population in front of the reader gets caught before it drives a decision.
Frequently asked questions
What did the ACR say about BI-RADS and pediatric patients?
In October 2026, the American College of Radiology published guidance, led by Stamatia Destounis, MD, stating that BI-RADS assessment categories and their associated malignancy probabilities were developed and validated for adult screening populations, not for diagnostic evaluation of a palpable breast mass in a child or adolescent. The ACR said management of pediatric breast masses should be based on clinical presentation and pediatric-specific imaging features, not on BI-RADS category-driven biopsy thresholds.
Why do BI-RADS categories overestimate cancer risk in children?
BI-RADS malignancy probabilities are calibrated to adult breast-cancer base rates. Pediatric breast cancer is exceedingly rare, with an incidence of about 0.08 newly diagnosed cases per 100,000 people, so applying the same category-to-risk mapping to a population with a near-zero baseline inflates the apparent probability of malignancy and pushes otherwise benign findings into higher-risk categories than the underlying biology supports.
What does the research show about BI-RADS 4 in pediatric patients?
A multicenter study of 227 patients younger than 21 who underwent pre-operative ultrasound found that 84% were assigned a BI-RADS category, and BI-RADS 4 — a category that by adult definition carries a 2% to 95% likelihood of malignancy — was the single most frequent assignment, given to 55% of the cohort. On pathology, every lesion in the study was benign. The authors, writing in the Journal of Pediatric Surgery in 2021, concluded that adult BI-RADS categories and their biopsy recommendations should not be routinely applied to pediatric patients.
What does this mean for structured reporting and AI in radiology?
The BI-RADS-in-children mismatch is a case study in a broader risk: any structured-reporting schema, including an AI system's risk categories and templates, encodes assumptions from the population it was built and validated on. Applying that same template unchanged to a different subpopulation, whether defined by age, anatomy, or clinical context, can systematically misclassify risk. Reporting systems, human or AI-assisted, need criteria that adapt to the population being read rather than one-size-fits-all adult-validated thresholds.
Source: American College of Radiology guidance on misuse of BI-RADS assessment categories on pediatric and adolescent breast masses (Destounis et al., October 2026), as reported by AuntMinnie and Newswise; underlying data from Davis et al., Journal of Pediatric Surgery (2021). Figures are rounded as reported.