Wednesday, December 3 | 3:20 p.m.-3:30 p.m. | SSPD04-3 | Room E350
Researchers are looking for safe ways to repurpose adult radiology AI models across pediatric imaging, but we're not there yet, according to this afternoon presentation.
Medical student Rachel Ivy of St. Jude Children's Research Hospital -- recognized for an RSNA 2025 Trainee Research Prize -- will share findings from a systematic review of scientific literature evaluating how adult-trained radiology AI models perform when tested in children. Out of over 2,000 articles identified, only 15 met inclusion criteria for full-text review.
The studies covered a range of AI tasks, including segmentation, object detection, and classification. A major finding was that only two models (both organ segmentation tasks on CT) showed statistically similar performance between adult and pediatric datasets, the group found.
"In nearly all cases, adult-trained models performed worse on pediatric data compared to adults," Ivy and colleagues wrote for their research abstract.
In addition, the degree of AI performance drop from adults to children varied widely. Some were modest drops, but some were substantial, according to the group.
Furthermore, six studies tested fine-tuning adult-trained AI models using limited pediatric data. While five of those studies showed statistically significant improvements, none exceeded the original performance on adults, Ivy and colleagues found.
Fine-tuning AI models with limited pediatric data can improve performance, but it does not fully close the gap, the group noted, adding that their findings underscore the need for dedicated pediatric imaging datasets.
This is an important issue pediatric radiologists and AI developers won't want to miss.
