AI and the World's Hard Problems · Medicine & drug discovery · Nature (computational biology) · 2026-10-23
FusionSwinUNETR: A multimodal deep learning model for abdominal mass segmentation in pediatric CT
Pediatric solid tumor research is limited by scarce data and by reliance on imaging alone. The team developed FusionSwinUNETR, which feeds patient metadata and clinician-written findings into a SwinUNETRv2 backbone alongside CT images to segment abdominal masses. Across 750 children with 12 mass types, it outperformed an image-only baseline on every metric, with Dice rising from 0.882 to 0.900 and boundary error (HD95) falling 46%, from 8.93 mm to 4.79 mm.