• Organization: Pediatric Brain Tumor Segmentation (BraTS-PEDs 2025)
  • Location: Los Angeles, CA
  • Period: Jul 2025 - Present

What problem we solved

Pediatric brain tumors require precise MRI segmentation for treatment planning, but class imbalance and noisy annotations make this difficult.

What I built

  • Engineered end-to-end preprocessing for 1000+ MRI volumes with TorchIO, SimpleITK, Nyul normalization, and N4 bias correction
  • Achieved greater than 95% consistency across scans
  • Architected GPU-efficient 3D models (DeepDenseTrans3D, GIETNet, ARES-UNet) with attention mechanisms

Measurable impact

  • Boosted Dice score by 8% over baseline U-Net
  • Composite loss (AFTL + Boundary + IoU + BCE) reduced false negatives by 15%
  • AMP and DDP optimization reduced training runtime by 25%

What I learned

Preprocessing and loss design can matter as much as architecture in medical imaging.