Repurposing adult AI radiograph models for neonatal care through continual learning: an international multicentre study

Despite growing clinical adoption of artificial intelligence in radiology, fewer than one in five commercially available models are cleared for paediatric use. Existing systems are trained predominantly on adult data and generalize poorly to neonates, yet efficient strategies for repurposing them without data sharing or large-scale local data curation are lacking. Here we present an international retrospective study of neonatal chest radiographs from 30 neonatal intensive care units across 20 countries and terri…

By the Truvace deskOctober 14, 20262 min read1 sourcePeer-reviewedTRV-2026-1366· cite & verify

Repurposing adult AI radiograph models for neonatal care through continual learning: an international multicentre study — "Governor Wolf Tours Bryn Mawr Hospital’s New Patient Pavilion and Neonatal Intensive Care Unit" by governortomwolf is licensed under CC BY 2.0. To view a copy of this license, visit https://creativecommons.org/licenses/by/2.0/.

In brief

Researchers took a deep learning model built to assess endotracheal tube position on adult X-rays and adapted it for newborns. They used chest radiographs from 30 neonatal intensive care units in 20 countries and territories, with each hospital supplying 50 images for sequential retraining rather than sending data to a central pool.

The work targets a gap in paediatric imaging AI, where few commercial tools are cleared for children and adult-trained systems perform poorly on neonates. The approach may offer a route to adapt existing models while keeping patient data local, but the study is retrospective and reports technical accuracy rather than outcomes for patients.

Main points

Fewer than one in five commercially available radiology AI models are cleared for paediatric use.
The study was an international retrospective analysis of neonatal chest radiographs from 30 NICUs in 20 countries and territories.
Each hospital contributed only 50 radiographs, and the model was retrained sequentially without centralized data pooling.
Continual learning gave 0.67 cm landmark localization error, versus 1.29 cm for hospital-specific fine-tuning and 4.30 cm for adult-only models.
Deep intubation detection improved at 27 of 30 sites.

The gain

Retraining an adult endotracheal tube model sequentially on 50 radiographs per hospital cut neonatal landmark localization error well below fine-tuning or adult-only use, without centralizing patient images.

The rundown

The starting point was a model for adult endotracheal tube assessment. The authors note that existing systems are trained mostly on adult data and generalize poorly to neonates, and that efficient ways to repurpose them without data sharing or large local data curation have been lacking.

Continual learning was tested as a privacy-preserving way to adapt the model. The hospitals contributed small local datasets of 50 radiographs each, and the model was updated in sequence across sites rather than by pooling images centrally.

Compared with hospital-specific fine-tuning and with the unmodified adult model, continual learning had lower anatomical landmark localization error at 29 of 30 sites. It also improved detection of deep intubations at 27 of 30 sites, and the authors conclude that collaborative privacy-preserving learning can address gaps in paediatric AI deployment.