AI-powered segmentation and prognosis with missing MRI in pediatric brain tumors.
Authors
Affiliations (13)
Affiliations (13)
- University of Pennsylvania, Department of Bioengineering, Philadelphia, PA, USA.
- Center for Data-Driven Discovery in Biomedicine (D3b), Children's Hospital of Philadelphia, Philadelphia, PA, USA.
- Department of Neurosurgery, Children's Hospital of Philadelphia, Philadelphia, PA, USA.
- Department of Radiology, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA.
- Department of Pediatrics, University of Michigan Medical School, Ann Arbor, MI, USA.
- Department of Neurological Surgery and Department of Pediatrics, University of California San Francisco, San Francisco, CA, USA.
- Division of Oncology, Children's Hospital of Philadelphia, Philadelphia, PA, USA.
- Department of Neurology and Pediatrics, University of California San Francisco, San Francisco, CA, USA.
- Division of Radiology, Children's Hospital of Philadelphia, Philadelphia, PA, USA.
- AI2D Center for AI and Data Science for Integrated Diagnostics, University of Pennsylvania, Philadelphia, PA, USA.
- Center for Data-Driven Discovery in Biomedicine (D3b), Children's Hospital of Philadelphia, Philadelphia, PA, USA. [email protected].
- Department of Neurosurgery, Children's Hospital of Philadelphia, Philadelphia, PA, USA. [email protected].
- Department of Neurosurgery, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA. [email protected].
Abstract
Brain MRI is the primary imaging modality for pediatric brain tumors, yet incomplete acquisitions are common, hindering the clinical utility of existing deep learning models for tumor segmentation and prognosis. These models are typically trained on complete MRI sequences and exhibit performance degradation when MRI sequences are missing at test time. In this retrospective study of 715 patients from the Children's Brain Tumor Network and BraTS-PEDs, and 43 patients with 157 longitudinal MRIs from PNOC003/007 clinical trials, we developed strategies for handling missing sequences. Methods included a dropout-trained segmentation model that randomly omitted FLAIR and/or T1w inputs during training, a generative model for image synthesis, copy-substitution heuristics, and zeroed inputs. The dropout model achieved robust segmentation under missing MRI, with ≤0.04 Dice drop relative to complete-input and stable prognostic accuracy in survival analysis using model-derived tumor volumes and clinical covariates. Generative synthesis achieved high image quality (SSIM > 0.90) and removed artifacts, benefiting visual interpretability. Together, these approaches can facilitate broader deployment of AI tools in real-world pediatric neuro-oncology settings.