Using a Radiologically Informed, Deep Learning Cascade to Refine Segmentations of Pediatric Brain Tumors from MRI

Automated brain tumor segmentation approaches, promoted by challenges such as the Brain Tumor Segmentation (BraTS) Challenge present a key opportunity to improve clinical practice for these patients. Being able to accurately and reliably segment and therefore monitor high grade glioma in pediatric brain tumor patients is a key example of this and is the target of the BraTS-PEDs 2025 challenge.

The current study presents a Radiologically informed, Deep Learning Cascade model, based on the residual encoder variant of nnU-Net, submitted to the BraTS-PEDs 2025 challenge. The goal of this model is to segment the subregions of brain tumors from coarse to fine and refining initial predictions through the two levels of the cascade. Our cascade model had some success in the previous BraTS-PEDs 2024 challenge, and the current submission proposes specific adaptations to further promote refinement by the model through the levels of the cascade, and to handle the challenges raised by tumor subregions that are underrepresented in training data due to them not being part of the radiological presentation in all HGGs.

Our novel adaptations of the cascade model provide robust segmentations for the BraTS-PEDs 2025 challenge validation data, achieving mean Dice scores of 0.692, 0.898, 0.699, and 0.945, and HD95 of 66.1, 6.2, 87.5, and 20.5 for the ET, NET, CC and ED, respectively.