Researchers involving the DQBM Menze Lab developed a deep learning model for automated segmentation of pediatric brain tumors. Trained on MRI data from 174 patients at University Children’s Hospital Zurich, the model accurately identified whole tumors and T2-hyperintense regions across diverse tumor types and locations.
The framework achieved performance comparable to human annotators and reduced manual contouring time by up to 83%. A streamlined combination of T1, contrast-enhanced T1 and T2 imaging also produced results close to the full four-sequence protocol. The publicly released models could support reproducible volumetric assessment and more efficient pediatric neuro-oncology workflows, although expert review remains necessary.